WO2025210350A1

WO2025210350A1 - Method for determining the presence or risk of disease and/or vascular state in a subject

Info

Publication number: WO2025210350A1
Application number: PCT/GB2025/050697
Authority: WO
Inventors: Charalambos Antoniades; Kenneth Chan; Pete TOMLINS; Matthew Kelly
Original assignee: Caristo Diagnostics Ltd; Oxford University Innovation Ltd
Current assignee: Caristo Diagnostics Ltd; Oxford University Innovation Ltd
Priority date: 2024-04-02
Filing date: 2025-04-02
Publication date: 2025-10-09
Anticipated expiration: 2026-10-02
Also published as: GB202404701D0; GB2640131A

Abstract

The invention relates to computer implemented methods for determining the presence or risk of disease and/or vascular state in a subject. The method comprises (a) using data gathered from a computer tomography (CT) scan along a length of two or more blood vessels to determine: (i) a value for the radiodensity and/or one or more radiomic features of the perivascular tissue surrounding each blood vessel or segment thereof; (b) using the value for the radiodensity and/or one or more radiomic features of the perivascular tissue surrounding each blood vessel or segment thereof to determine the presence of disease and/or vascular state in each blood vessel, or segment thereof; and (c) generating an output value based on the number of vessels or segments thereof with disease and/or vascular state that indicates the subject's presence or risk of disease and/or vascular state. The invention also relates to computer program products, systems and machine learning or statistical modeals for performing the methods of the invention. Also disclosed is methods of training said machine learning or statistical models, and related systems.

Description

METHOD FIELD OF THE INVENTION The present invention relates to computer implemented methods for determining the presence or risk of disease and/or vascular state in two or more blood vessels _{of a subject. The methods may be used to predict cardiac mortality risk, risk of a} patient suffering a cardiovascular event, determine disease development, progression or regression, monitor the development of cardiovascular disease, guide pharmacological treatment decisions, and monitor responses to medical treatments. The invention also provides a computer program product configured to carry out these methods, and to a system comprising one or more processors. BACKGROUND _{Atherosclerosis is a progressive process in which an artery wall thickens as a} result of invasion and accumulation of white blood cells. This inflammatory process results in the development of plaques within the vessel wall containing living white blood cells, dead cell debris and fatty deposits including cholesterol and triglycerides. Stable atherosclerotic plaques, which tend to be asymptomatic, are typically rich in extracellular matrix and smooth muscle cells, while unstable plaques are rich in macrophages and foam cells. The extracellular matrix in unstable plaques separating the lesion from the arterial lumen (also known as the fibrous cap) is usually weak and prone to rupture. Ruptures of the fibrous cap eventually induce clot formation in the lumen, and such clots can block arteries or detach, move into the circulation and eventually block smaller downstream vessels causing thromboembolism. Chronically expanding plaques are frequently asymptomatic until vessel occlusion (stenosis) is severe enough that blood supply to downstream tissue is insufficient. _{Atherosclerosis is often asymptomatic for decades because the arteries can} enlarge at plaque locations and blood flow is not immediately affected. Indeed, plaque ruptures are also asymptomatic unless they result in sufficient narrowing or closure of an artery that impedes blood flow to different organs so as to induce symptoms. Typically, the disease is only diagnosed when the patient experiences other cardiovascular disorders such as stroke or heart attack. Atherosclerosis may cause narrowing in the coronary arteries, which are responsible for bringing oxygenated blood to the heart, and this can produce symptoms such as the chest pain of angina, shortness of breath, sweating, nausea, dizziness or light-headedness, breathlessness or palpitations. Cardiac arrhythmias may also result from cardiac ischemia. Atherosclerosis that causes narrowing in the carotid arteries, which supply blood to the brain and neck, can produce symptoms such as a feeling of weakness, not being able to think straight, difficulty speaking, becoming dizzy and difficulty in walking or standing up straight, blurred vision, numbness of the face, arms, and legs, severe headache and losing consciousness. These symptoms may also be present in stroke, which is caused by marked narrowing or closure of arteries going to the brain leading to brain ischemia and death of cells in the brain. Peripheral arteries, which supply blood to the legs, arms, and pelvis may also be affected. Symptoms can include numbness within the affected limbs, as well as pain. Plaque formation may also occur in the renal arteries, which supply blood to the kidneys. Plaque occurrence and accumulation leads to decreased kidney blood flow and chronic kidney disease, which, like all other areas, are typically asymptomatic until late stages. Vascular inflammation is a key feature in atherogenesis and plays a critical role in atherosclerotic plaque stability by triggering plaque rupture leading to acute coronary syndromes (see Ross R. N Engl J Med 1999;340:115-26, and Major AS _{et al Circulation 2011;124:2809-11). Importantly, more than 50% of acute} coronary syndromes are caused by highly inflamed but anatomically non- significant atherosclerotic plaques (Fishbein MC et al. Circulation 1996;94:2662- 6). Cardiac computed tomography angiography (CCTA) can quantify the extent, distribution and characteristics of coronary plaques, but this is not sufficient for optimal risk prediction in individuals. It is well known that most acute myocardial infarctions (MIs) occur secondary to occlusion in vessels with minor coronary plaque disease that erodes or ruptures. This relates to the biology of the underlying coronary plaque, particularly inflammation. Coronary computed tomography angiography (CCTA) is the first line investigation for chest pain, and it is used to guide revascularisation. However, the widespread adoption of CCTA has revealed a large group of individuals without obstructive coronary artery disease (CAD), with unclear prognosis and management. These subjects are often reassured and discharged without specific treatment or follow up, as their management and outcome are unclear. Measurement of coronary inflammation from CCTA using perivascular fat attenuation index (FAI) Score could enable cardiovascular risk prediction and guide the management of individuals without obstructive CAD. _{In the PROMISE trial, 54% of adverse events occurred in patients without} significant stenoses, whereas patients with significant stenoses accounted for only 12% of the population undergoing CCTA. Thus, more than half of the aggregate risk of adverse cardiovascular events is not identified by coronary stenoses in people who undergo CCTA. This limitation is a driver of ‘residual risk’ that results in adverse cardiovascular outcomes, despite efforts to manage cardiovascular disease (CVD) according to current recommendations. CCTA can identify patients with plaque characteristics associated with high risk, such as low-attenuation plaque, napkin ring sign, positive remodelling and spotty plaque calcification; however, these provide only modest incremental information in individual patients. The predictive value of high-risk plaque (HRP) features was studied in both the PROMISE and SCOT-HEART trials. In the PROMISE Study, approximately 15% of patients were found to have HRP on their CCTA and HRP _{was associated with more adverse cardiovascular events, although the major} adverse cardiovascular events (MACE) endpoint included revascularisation which may not reflect the additional value of HRP above and beyond stenosis severity. Nevertheless, most patients with HRP did not have cardiovascular events, whereas many patients without HRP did, indicating the limited predictive value of HRP on CCTA. Indeed, of the 1019 HRPs identified on CCTA, only 24 subsequent non-fatal MIs occurred, demonstrating that the absolute risk of a cardiovascular event in relation to a single plaque, identified at a single time point, is extremely low. This observation is consistent with other CCTA and intravascular ultrasound (IVUS) studies. In the SCOT-HEART trial, 1376 HRP features on CCTA were detected in 608 of 1769 participants. The likelihood of an adverse cardiac event during follow-up was increased in the subjects with HRP features, but the absolute increase in risk was very small (4.1% with HRP vs 1.4% without HRP). Importantly, more than one-third of the events occurred in subjects without HRP features. A more detailed analysis identified low-attenuation non-calcified plaque burden as the most specific HRP feature predictive of adverse events. Furthermore, recent discoveries highlight the importance of cellular inflammatory mechanisms in the vascular wall as drivers of disease progression and risk of events. Coronary artery inflammation is a major factor in CAD progression, and a key determinant of high-risk plaques that drive adverse clinical events, in addition to the contributions of stenosis, flow limitation or high-risk plaque features. Imaging perivascular adipose tissue (PVAT) around the coronary arteries (pericoronary adipose tissue; PCAT) has emerged as a promising technique to image inflammation in the coronary artery wall. A key recent discovery is that PVAT ‘senses’ the presence of inflammation in the wall of the coronary artery. These signals transduce changes in PVAT differentiation, leading to smaller, less lipid- rich adipocytes, greater inflammatory cell infiltration and higher tissue water content. These changes modify tissue attenuation values in a 3D distribution around the coronary artery that can be detected using cardiac computed tomography angiography (CCTA). Simple measures of PVAT attenuation require corrections for anatomical, technical factors, and patient and clinical variables. However, the validity of measuring changes in PCAT attenuation has been reproduced in research studies performed using similar techniques in different patient groups. Goeller et al. (JAMA Cardiol. 2018; 3:858) reported that uncorrected PCAT CT attenuation was increased around culprit lesions compared to nonculprit lesions of patients with ACS and the lesions of matched controls. In a multivariate _{analysis, low- and intermediate-attenuation non-calcified plaque burden and} PCAT CT attenuation were independently associated with the presence of culprit lesions. They concluded that combined quantitative high-risk plaque features and PCAT CT attenuation may allow for a more reliable identification of vulnerable plaques. The recent ORFAN (Oxford Risk Factors and Non-invasive imaging) study by the present inventors validated the efficacy of fat attenuation index (FAI) scores, and evaluated its real-world impact on patient management. The study revealed that cardiac death or MACE occurred in 3.4% or 8.7% of patients, respectively, without obstructive CAD. In patients with either obstructive or non-obstructive disease, those with a FAI score in the LAD above the 75th percentile, as compared to the _{25th percentile, had 20.2- or 6.7-times higher risk of cardiac death or MACE,} respectively. Among those without obstructive CAD, the risk of cardiac death or _{MACE was still 10.5- or 4.8-times higher, respectively. An artificial intelligence-} assisted prognostic model (AI-risk), that included FAI-Score-based inflammatory risk, significantly reclassified the clinical risk of patients compared to a clinical risk factors-based prediction model (QRISK3) in the whole cohort and those without obstructive CAD, resulting in change of treatment for a significant proportion of _{patients, including statin initiation, statin-dose intensification and/or additional} treatments, such as colchicine. The study concluded that patients with high coronary inflammation, measured by FAI Score, have substantially higher risk for _{cardiac death and MACE, and the AI-Risk leads to reclassification and change of} management in a substantial proportion of patients undergoing routine CCTA. Accordingly, this CCTA risk score can be used as a precision medicine tool. Although such CCTA approaches have been proven as a promising development in recent years in the monitoring and treatment of coronary disease, there remains a need to develop more precise and reliable methods for generating a risk score for predicting cardiovascular events. In particular, there is a need to enable the earlier detection of cardiovascular risk, and minimise false negative results obtained through previously known and less-sensitive CCTA functional biomarkers. SUMMARY OF INVENTION In a first aspect, the invention provides a computer implemented method for determining the presence or risk of disease and/or vascular state in a subject. The method comprises the steps of: (_{a) using data gathered from a computer tomography (CT) scan along a length} of two or more blood vessels to determine: (_{i) a value for the radiodensity and/or one or more radiomic features of} _{the perivascular tissue surrounding each blood vessel or segment} thereof; (_{b) using the value for the radiodensity and/or one or more radiomic features} _{of the perivascular tissue surrounding each blood vessel or segment} _{thereof to determine the presence of disease and/or vascular state in a} _{subject in each blood vessel or segment thereof; and} _{(c) generating an output value based on the number of vessels or segments} _{thereof with disease and/or vascular state, that indicates the subject’s} presence or risk of disease and/or vascular state. Steps (a)(i), (b) and (c) preferably comprise determining a value for the radiodensity and/or one or more radiomic features of the perivascular tissue surrounding each blood vessel, using this value to determine the presence of disease and/or vascular state, and generating an output value based on the number of vessels having disease and/or (adverse) vascular state. Additionally, or alternatively, steps (a)(i), (b) and (c) comprise analysing blood vessel segments. In a preferred embodiment of the invention, step (b) comprises determining the _{presence of inflammation in each blood vessel or segment thereof; and preferably,} step (b) comprises determining an output value based on the number of inflamed _{blood vessels or segments thereof.} Preferably, the term “segment” in the context of the invention relates to segments _{encompassing a length of the blood vessel being analysed. In other words,} preferably the segments comprise a length of the centreline. In a preferred embodiment, the blood vessels are coronary arteries. The value for the radiodensity is preferably taken from the perivascular adipose tissue _{surrounding the blood vessels or segments thereof.} In a second aspect, the invention provides a computer program product comprising executable instructions which, when executed by one or more processors, cause the one or more processors to perform the method of the first aspect. In a third aspect, the invention provides a system comprising one or more processors configured to perform the method of the first aspect. _{According to a fourth aspect, the invention provides a computer-implemented} method for determining the presence or risk of disease and/or vascular state in a subject, comprising: a_{ccessing a metric indicative of the number of inflamed blood vessels, or} _{segments thereof having disease and/or vascular state, from two or more blood} vessels of the subject; and i_{nputting the metric indicative of the number of blood vessels or segments} _{thereof having disease and/or vascular state, into a trained machine learning or} statistical model to determine the presence or risk of disease and/or vascular state. _{The invention also provides, in a fifth aspect, a (e.g. computer implemented)} method of training a machine learning or statistical model for determining the presence or risk of disease and/or vascular state in a subject, comprising: generating a reference dataset that comprises: (i) a plurality of reference metrics indicative of the number of b_{lood vessels, or segments thereof having disease and/or vascular state,} _{from two or more blood vessels of the subject; and} _{(ii) a plurality of reference classifications of a cardiac mortality} risk or risk of a subject suffering a cardiovascular event associated with t_{he reference metrics indicative of the number of blood vessels or} _{segments thereof having disease and/or vascular state; and} training the machine learning or statistical model using the reference dataset to learn a relationship between the metrics indicative of the number of _{blood vessels or segments thereof having disease and/or vascular state, and the} classifications of the presence or risk of disease and/or vascular state in a subject. According to a sixth aspect, the invention provides a machine learning or statistical model trained using the method of the fifth aspect. In a seventh aspect, the invention provides a system for determining the presence or risk of disease and/or vascular state in a subject, comprising at least one processor in communication with at least one memory device, the at least one memory device having stored thereon instructions for causing the at least one processor to perform a method according to the fourth or fifth aspect of the invention. Further disclosed is a computer program product comprising executable instructions which, when executed by one or more processors, cause the one or more processors to perform the method of the fourth or fifth aspect. The methods according to the invention can be used to guide pharmacological treatment decisions, monitor responses to medical treatments by assessing dynamic changes in coronary inflammation; stratify subjects according to their risk of cardiac mortality or risk of suffering a cardiovascular event; quantify vascular inflammation, fibrosis, oedema and vascularity; predict disease development, progression or regression; and measure a change in disease or vascular status. It is preferable that the output value be corrected for biological and technical factors, to improve accuracy and sensitivity. It is preferred that these corrections be carried out by a processing system i.e. be computer implemented. BRIEF DESCRIPTION OF DRAWINGS Figure 1 shows the additive prognostic value of high coronary inflammation recorded in 1, 2 or 3 epicardial arteries. Prognostic value for cardiac mortality in the whole population (a), patients without obstructive coronary artery disease (CAD) (B) or with obstructive CAD (C). Similarly, the prognostic value for MACE in the whole population (D), patients without obstructive coronary artery disease (CAD) (E) or with obstructive CAD (F). Inflamed coronary artery defined as having FAI-Score >75th percentile. Reference: all 3 coronary arteries (LAD, LCx and RCA) with FAI Score <25th percentile. Figure 2 illustrates the three main coronary arteries, with the two shaded vessels designated as inflamed, and the un-shaded vessel designated as non-inflamed. Figure 3 schematically illustrates an example of a system suitable for implementing embodiments of the method. The system 1100 comprises at least one server 1110 which is in communication with a reference data store 1120. The server may also be in communication with other hardware or system 1130 which may be operated by a healthcare professional, for example over a communications network 1140. Figure 4 describes an example of a suitable server 1110 which may be used to _{implement the methods of the invention. In this example, the server includes at} _{least one microprocessor 1200, a memory 1201, an optional input/output device} _{1202, such as a keyboard and/or display, and an external interface 1203,} _{interconnected via a bus 1204 as shown. In this example the external interface} _{1203 can be utilised for connecting the server 1110 to peripheral devices, such as} _{the communications networks 1140, reference data store 1120, other storage} devices, or the like. _{Figure 5 illustrates numbered segments of the coronary arteries. Along the right} coronary artery (RCA) there is a proximal segment 1, medial segment 2, distal _{segment 3, and posterior descending artery (R-PDA) 4. Along the left main} _{coronary artery 5 there is the left anterior descending artery LAD comprising the} proximal segment 6, medial segment 7, apical segment 8, first diagonal branch _{(D1) 9, and second diagonal branch (D2) 10. Along the left circumflex artery (Cx),} _{there is the proximal segment 11, obtuse marginal branch (OM) 12, distal} _{segment 13, left posterolateral branch (PL) 14, and posterior descending artery} _{(PDA) 15.} _{Figure 6 schematically depicts the Study design and data flow of the clinical study} of the Examples. Abbreviations: HES, Hospital episodes statistics; NHS, National Health Service; NICOR, National Institute for Cardiovascular Outcomes Research; ONS, Office of National Statistics. Figure 7 shows the cardiovascular risk prediction in the presence or absence of _{obstructive CAD. Forrest plot showing hazard ratios for individual clinical} outcomes and MACE (cardiac mortality, myocardial infarction, new heart failure) over a period of 10 years after the CCTA in 40,091 Cohort A patients. CAD=coronary artery disease. HR adjusted for age, sex, cardiovascular risk factors (diabetes, hypertension, hyperlipidaemia, smoking status), medications (betablockers, calcium channel blockers, nitrates, statins, angiotensin-converting enzyme inhibitors, antiplatelets and direct oral anticoagulants), past myocardial _{infarction and history of revascularisation (PCI or CABG). Abbreviations: CAD,} coronary artery disease; MACE, Major adverse cardiac events. Figure 8 shows the independent calibration of AI-Risk and the AI-Risk Classification system, in predicting cardiovascular events in cohort B. Calibration curves of the AI-Risk (as a continuous variable) in the UK (external validation) cohort, across the whole population (A), in those with no obstructive (B) or obstructive (C) CAD. Calibration curves of the AI-Risk Classification categories is presented across the whole population (D), as well as in those with no obstructive (E) or obstructive CAD (F). _{Figure 9 illustrates the ccalibration of AI-Risk against observed cardiac mortality} among individuals with obstructive coronary artery disease (CAD), after adjusting for calcified plaque volume (CP) and non-calcified plaque volume (NCP). (A) There was minimal improvement in the performance of CP/NCP plaque-adjusted AI-Risk model compared to the original AI-Risk model. (B) Adjustment of the AI- Risk classification system for CP and NCP volumes had minimal impact on its performance in any of the three risk categories. _{Figure 10 relates to the AI-Risk Classification and cardiovascular risk prediction.} _{Kaplan-Meier (KM) curves for the ability of AI-Risk Classification to prediction} cardiac mortality in (A) the whole cohort, (B) patients with no obstructive CAD, (C) patients with obstructive CAD. KM curves for prediction of MACE using the same classification are presented for (D) the whole cohort, (E) patients with no obstructive CAD, (F) patients with obstructive CAD. AI, artificial intelligence; CAD, coronary artery disease. _{Figure 11 illustrates the prognostic value of FAI Score and AI-Risk in patients with} _{no or minimal coronary atherosclerosis. Kaplan-Meier curves for predicting} cardiac mortality with FAI Score in the (A) LAD, (B) LCX, (C) RCA, and (D) AI-Risk classification. Kaplan-Meier curves for predicting MACE with FAI Score in the (E) LAD, (F) LCX, (G) RCA, and (H) AI-Risk classification. Figure 12 shows the incremental discriminatory value of the AI-Risk Classification _{system above a risk factors-based model. Incremental discriminatory value of the} AI-Risk Classification system for cardiac mortality above the QRISK3 classification. In the whole cohort of Cohort B (A) and those with no obstructive CAD (B). Similar comparisons are presented for MACE (composite of cardiac mortality, non-fatal myocardial infarction and new onset of heart failure) (panels C and D). The thick black line represents events, whereas the thin black line represents non-events. The difference between the black dots represents the continuous NRI, the difference between the grey dots represents the median improvement, whereas the shaded area reflects the IDI. Reclassification table for discrete cardiac risk groups in the two study cohorts comparing AI-Risk Classification against QRISK3 classification risk categories. Risk categories based on QRISK3 include: low/mid risk (<10% 10y risk for MACE), high risk (10- 19%) and very high risk (≥20%). The respective risk categories for AI-Risk Classification include: Low/mid risk category (AI-Risk <5% and FAI Score <75th percentile in the LAD/RCA and <95th percentile in the LCX); High risk category (AI-Risk 5% to <10% or FAI Score in the LAD/RCA between 75th and 90th percentile or FAI Score in the LCx >95th percentile), and Very high risk (AI-Risk ≥10% or FAI Score at LAD/RCA >90th percentile). DETAILED DESCRIPTION _{The inventors have developed computer implemented methods which are able to} _{provide valuable clinical information on the risk of major adverse cardiac events} _{(MACE) (e.g. cardiac mortality risk or risk of a subject suffering a cardiovascular} _{event), and other disease or vascular states, based on the number of blood} _{vessels or segments thereof with disease and/or adverse vascular states, of a} _{given subject. For example, the disease or vascular state of each vessel, or} _{segment thereof, may be selected from: inflammation, oedema, fibrosis,} _{vascularity lipolysis and/or adipogenesis; of a given subject. The output of the} methods may be used on their own, or in combination with other established _{clinical risk factors to provide a more reliable, clinically valuable risk score, or} clinical picture of a subject’s vascular pathology. This targeted approach can lead to more accurate and precise diagnosis, and enable clinicians to make more informed decisions regarding patient care and treatment options. The subject may be an individual who has been diagnosed as suffering from a condition associated with vascular inflammation, oedema, fibrosis, vascularity, lipolysis and adipogenesis; or who is suspected of, or at risk of, suffering from such conditions, in particular vascular inflammation affecting the coronary vessels. Alternatively, the patient may be a healthy individual who has not been diagnosed _{as suffering from a condition associated with vascular inflammation, oedema,} fibrosis, vascularity, lipolysis and/or adipogenesis, and/or who is not known to be at risk of suffering from such conditions. _{The risk score of the invention can be used as an adjunctive tool in routine clinical} CT angiograms to identify patients at high risk of cardiac events and mortality including a sensitive and specific screening tool in people who are apparently healthy and low-risk according to the traditional interpretation of their scans. Thus, _{the methods as claimed find utility both in primary prevention (healthy population} with no diagnosis of heart disease yet) and secondary prevention (patients with a diagnosis of coronary artery disease), to identify an individual’s risk status beyond traditional risk factors, to guide pharmacological treatment decisions, predict risk of disease development, progression or regression; and to monitor response to appropriate medical treatments. In the first aspect of the invention, the computer implemented method comprises the step of (_{a) using data gathered from a computer tomography (CT) scan along a length} of two or more blood vessels to determine; (i) a value for the radiodensity and/or one or more radiomic features of the p_{erivascular tissue surrounding each blood vessel or segment thereof;} _{(b) using the value for the radiodensity and/or one or more radiomic features} of the perivascular tissue surrounding each blood vessel or segment t_{hereof to determine the presence of disease and/or vascular state (e.g.} _{adverse vascular state) and} _{(c) generating an output value based on the number of vessels or segments} _{thereof with disease and/or vascular state (e.g. adverse vascular state),} that indicates the subject’s risk of cardiac mortality or risk of suffering a cardiovascular event. The disease and/or vascular state may be selected from one or more of: inflammation, oedema, fibrosis, vascularity, lipolysis and adipogenesis; preferably inflammation. Preferably, step (b) comprises determining the presence of inflammation in each blood vessel. Therefore in step (c), the number of inflamed vessels, or segments _{thereof, is used to generate the output value.} In another preferred embodiment, step (b) of the method comprises comparing the value of an inflammation, oedema, fibrosis, vascularity metric derived from the radiodensity value and/or one or more radiomic features (preferably radiodensity value) to a pre-determined cut-off value or using the absolute value to determine _{whether each blood vessel or segment thereof has inflammation, oedema,} fibrosis, vascularity, lipolysis and/or adipogenesis. It is particularly preferred that the metric is an inflammation metric e.g. it may comprise determining the fat attenuation index (FAI) or perivascular adipose tissue (FAIPVAT). Alternatively, step (b) may comprise determining one or more of: (_{i) calcium index (Calcium-i);} _{(ii) perivascular water index (PVWi);} _{(iii) fat attenuation index of perivascular adipose tissue (FAIPVAT);} _{(iv) fibrous plaque index (FPi);} _{(v) perivascular water index (PVWi);} _{(vi) volumetric perivascular characterisation index (VPCI);} _{In the most preferred embodiment, the two or more blood vessels are from a} _{particular group of vessels supplying blood to a major organ (such as the heart or} brain). For example, the two or more blood vessels are preferably the coronary arteries. Alternatively, the two or more blood vessels are the carotid arteries, or _{branches of the aorta. Such that, the number of vessels or segments thereof} which have disease and/or adverse vascular state (e.g. are inflamed, or have oedema, fibrosis, vascularity, lipolysis and/or adipogenesis) are the number of _{coronary arteries or number of carotid arteries (or segments thereof) that are} characterised as such. In a preferred embodiment, step (b) comprises determining the presence of _{disease and/or vascular state (preferably inflammation) in each of the coronary} arteries, carotid arteries, or segments/branches of the aorta, and step (c) comprises generating an output value based on the number of coronary arteries, _{carotid arteries or segments/branches of the aorta having a disease and/or} vascular state. By quantifying the amount of disease or adverse vascular state (e.g. inflammation) within these particular groups of vessels, a more accurate risk score can be generated versus previously known methods in which the highest FAI score of any given vessel was used in a model to predict cardiovascular risk. The group of coronary arteries which supply blood to the heart, for example, _{comprise three distinct main arteries – the right coronary artery (RCA), the left} anterior descending coronary artery (LAD) and the left circumflex artery (LCx). When the number of diseased (e.g. inflamed) vessels is the number of diseased (e.g. inflamed) coronary arteries, preferably the coronary arteries are selected from two or more of, preferably each of: the right coronary artery, left anterior _{descending artery and left circumflex artery.} The present inventors have identified that the number of coronary arteries that _{exhibit a diseased or adverse vascular state is strongly correlated with risk of} adverse cardiac events. For instance, subjects with zero inflamed coronary arteries have the lowest risk, whereas patients with 3 inflamed coronary arteries have the highest risk. If a subject has only one inflamed coronary artery, their chances of suffering a heart attack are lower than subjects having two or three inflamed coronary arteries. The determination of this number of arteries from this _{particular group of vessels can provide an enhanced prognostic tool over and} above previous methods which only took the highest FAI score of any given vessel into account when determining cardiovascular risk. Accordingly, the number of _{diseased (e.g. inflamed) vessels or segments can be taken into account as a} _{standalone biomarker of cardiovascular risk (e.g. risk of heart attack), or it can be} _{taken into account in combination with other established biomarkers and risk} _{factors to provide a more accurate risk prediction score of adverse cardiac events,} or indication of disease and/or vascular state. In a preferred embodiment, the number of coronary artery segments with disease _{and/or adverse vascular state is determined and used in step (c) to provide the} output value. Figure 5 illustrates the coronary artery vessel segments according to the AHA definition. Along the right coronary artery (RCA) there is a proximal segment 1, medial segment 2, distal segment 3, and posterior descending artery _{(R-PDA) 4. Along the left main coronary artery 5 there is the left anterior} _{descending artery LAD comprising the proximal segment 6, medial segment 7,} _{apical segment 8, first diagonal branch (D1) 9, and second diagonal branch} _{(D2) 10. Along the left circumflex artery (Cx), there is the proximal segment 11,} _{obtuse marginal branch (OM) 12, distal segment 13, left posterolateral branch} _{(PL) 14, and posterior descending artery (PDA) 15.} The number of diseased (preferably inflamed) coronary artery segments may be used in step (c) to indicate the presence or risk of disease or vascular state of a subject. For example, in one embodiment, the method can be used to predict cardiac mortality risk, or risk of a patient suffering a cardiovascular event. The carotid arteries is another group of vessels that supply blood to the brain, and comprise the left common carotid artery and right common carotid artery. These arteries travel from the upper chest to the skull, and each divides into two _{branches – the left and right internal carotid artery, and the left and right external} _{carotid arteries. A blockage or clot in one of the carotid arteries can impede blood} flow to the brain and cause a stroke. Consequently, by quantifying the number of blood vessels or segments thereof having disease and/or adverse vascular state from this group of vessels, an enhanced tool for the risk of predicting a subject suffering a stroke can be determined. In a preferred embodiment, the number of diseased and/or adverse vascular state _{(e.g. inflamed) carotid arteries are determined; and are selected from the left} common carotid artery and the right common carotid artery. In this embodiment, a _{subject may have zero, one or two carotid arteries characterised as having} disease and/or vascular state. Additionally or alternatively, the number of inflamed carotid artery segments with _{disease and/or adverse vascular state is determined. The carotid artery segments} _{may be selected from the left common carotid artery, the right common carotid} artery, the left internal carotid artery, the right internal carotid artery, the left _{external carotid artery and the right external carotid artery. Therefore, a subject} _{may have zero to six carotid artery segments with disease and/or adverse} vascular state. When the number of vessels or segments thereof in step (c) is the number of carotid arteries, and/or number of carotid artery segments having disease and/or adverse vascular state, the method can be used to predict risk of a subject _{suffering a stroke. The method can also be used for determining the risk of cardiac} mortality, or risk of a subject suffering a cardiovascular event. In another embodiment, the method involves determining a value for the _{radiodensity and/or one or more radiomics features of blood vessel segments} _{selected from segments or branches of the aorta. They may be selected from} _{thoracic and/or abdominal segments of the aorta.} The thoracic aorta can be further subdivided into the segments: ascending aorta, aortic arch, and descending aorta. Accordingly, in an embodiment, step (a)(i) _{involves determining a value for the radiodensity and/or one or more radiomics} features of the ascending aorta, aortic arch and descending aorta. These values are used in steps (b), and a corresponding output value is generated in step (c) based on the number of segments of the aorta with disease and/or (adverse) vascular state. _{Alternatively, the method may comprise determining a value for the radiodensity} _{or one or more radiomics features of segments or branches of the thoracic aorta} in step (a)(i). These may be selected from: bronchial arteries, pericardial arteries, _{intercostal arteries, superior phrenic arteries, oesophageal arteries and} mediastinal arteries. In another embodiment, the method comprises determining a value for the radiodensity or one or more radiomics features segments or branches of the _{abdominal aorta in step (a)(i). These may be selected from: inferior phrenic} arteries, adrenal arteries, celiac trunk, renal arteries, gonadal arteries, superior mesenteric artery, lumbar arteries, inferior mesenteric artery, common iliac _{arteries and median sacral artery.} In one embodiment of the invention, the method comprises determining the number of inflamed blood vessel segments and using this value in step (c) of claim 1. For example, this may include coronary segments (example of potential segmentation of the coronary arteries shown in Figure 5), carotid artery segments, thoracic aorta segments or abdominal aorta segments, as indicated above. For the avoidance of doubt, the methods of the invention utilise CT scan data that has been obtained in vivo, by scanning a living body, but the claimed methods are not practised on the living human or animal body. Consequently, the method of the invention is non-invasive and is based on the analysis of conventional CT images; it does not require any additional image acquisition. Radiodensity, which is measured in Hounsfield units (HU), is a measure of the relative inability of X-rays to pass through material. The term “rradiodensity” is synonymous with the term “attenuation” and the two terms can be used interchangeably. Measurement of attenuation values allows tissue types to be distinguished in CT on the basis of their different radio-opacities. Fat is not very radiodense, and it _{typically provides a much lower radiodensity than muscle, blood and bone. The} exact HU ranges which correspond to different tissue types typically vary depending on factors such as CT scan parameters, and the type of software used _{to reconstruct and analyse the medical imaging data.} For instance, the following software programs designate vascular and perivascular tissue types as follows. Coronary Plaque Analysis 2.0.3 syngo.via FRONTIER, Siemens ^_{Calcified plaque: >700HU} _{^ Lumen: 150 – 700 HU} _{^ Non calcified fibrotic plaque: 30 – 150 HU} _{^ Non calcified lipid rich: ≤ 30 HU} QAngioCT version 3.1.3.13 Medis Medical Imaging Systems ^_{Dense calcium: >351 HU} _{^ Fibrous plaque: 151-350 HU} _{^ Fibrofatty plaque: 31 – 150 HU} _{^ Necrotic core: -30 – 30 HU} SUREPlaque, version 6.3.2; Vital Images ^_{Calcified plaque: >150 HU} _{^ Fibrous plaque: 50 – 150 HU} _{^ Fatty plaque: -100 – 49 HU} A person skilled in the art of cardiology and medical imaging technology is therefore able to distinguish different tissue types from CT image data depending on the equipment and software used, without the need for an exact definition in HU. The term “voxel” has its usual meaning in the art and is a contraction of the words “volume” and “element” referring to each of an array of discrete elements of volume that constitute a notional three-dimensional space. _{Step (a)(i) may comprise determining a value for one or more radiomic features} from the CT data. The value for each radiomic feature is then included in step (b) and (c) to generate an output value that indicates the presence of disease and/or _{vascular state in each blood vessel or segment thereof.} In WO 2020/058713 A1, it was identified that the use of radiomic features adds incremental value beyond traditional risk factors and established CCTA risk classification tools in predicting future adverse cardiovascular events and evaluating cardiovascular health and risk, and further aids the detection of vascular inflammation, local plaque inflammation, and the presence of unstable coronary lesions. Preferably, a value for the radiodensity is determined in addition to one or more radiomic features, preferably two or more radiomic features. The use of two or more radiomic features provides a ‘radiomic signature’ and provides a tool for further characterising the blood vessel or segment thereof of interest. For embodiments in which one or more radiomic features are determined, the output value determined in step (b) may provide a measure of the texture of the vascular region. If an indication of texture is to be provided, at least one radiomic feature may provide a measure of the texture. If more than one radiomic feature has been determined, each of the radiomic features may provide a measure of the texture _{of the perivascular space surrounding the blood vessel (i.e. each of the at least} two radiomic features may be texture statistics). Alternatively, one or more radiomic feature may be required to provide a measure of the texture. In a preferred embodiment, the methods of the invention comprise determining two or more radiomic features. More preferably, three or more, or four or more radiomic features. If two or more radiomic features are determined of the radial segment, a radiomic signature may be calculated for the radial segment. The radiomic signature is described in detail in WO 2020/058713, the entirety of which is herein incorporated by reference. When deriving a radiomic signature, the method may comprise using a dataset, in _{particular a radiomic dataset, to construct a radiomic signature or score. This} score may then be used to characterise the pathology or state of the vascular region, characterise the inflammation of a plaque, provide a cardiovascular risk score, or guide pharmacological treatment. Accordingly, the radiomic features can be used as a tool to determine the presence of disease and/or vascular state in _{each blood vessel or segment thereof.} _{The radiomic signature may be calculated on the basis of a (second) plurality of} _{perivascular radiomic features. The dataset may comprise the measured values} of a (first) plurality of perivascular radiomic features of a perivascular region obtained from medical imaging data for each of a plurality of individuals. The plurality of individuals may comprise a first group of individuals having reached a clinical endpoint indicative of cardiovascular risk, and/or particular biological state such as inflammation, fibrosis, oedema, vascularity, lipolysis, adipogenesis or combinations thereof; and a second group of individuals having not reached a clinical endpoint indicative of cardiovascular risk and/or the particular biological state(s). The second plurality of radiomic features may be selected from amongst _{the first plurality of radiomic features, in particular to provide a radiomic signature} for predicting cardiovascular risk, as determined from or using the dataset, for example using a machine learning algorithm. The radiomic signature may therefore be calculated on the basis of further radiomic features (for example _{selected from the (first) plurality of radiomic features) in addition to the at least two} radiomic features. The radiomic features are preferably selected from one or more of, more preferably at least two of: Short Run High Gray Level Emphasis, High Gray Level Emphasis, High Gray Level Run Emphasis, Autocorrelation, Sum Average, Joint _{Average, and High Gray Level Zone Emphasis, Skewness, Skewness LLL,} Kurtosis, 90th Percentile, 90th Percentile LLL, Median LLL, Kurtosis LLL, and Median, Run Entropy, Dependence Entropy LLL, Dependence Entropy, Zone Entropy LLL, Run Entropy LLL, and Mean LLL, Small Area Low Gray Level Emphasis, Low Gray Level Zone Emphasis, Short Run Low Gray Level Emphasis, Low Gray Level Run Emphasis, Low Gray Level Emphasis, Small Dependence Low Gray Level Emphasis, Gray Level Variance LLL (GLSZM), Gray Level Variance (GLDM), Variance, Gray Level Variance (GLDM), Difference Variance LLL, Gray Level Variance LLL (GLRLM), Variance LLL, Gray Level Variance LLL (GLDM), Sum of Squares, Contrast LLL, Mean Absolute Deviation, Interquartile Range, Robust Mean Absolute Deviation, Long Run Low Gray Level Emphasis, Difference Variance, Gray Level Variance (GLSZM), Inverse Difference Moment Normalized, Mean Absolute Deviation LLL, Sum of Squares LLL, and Contrast, Zone Entropy, Gray Level Non Uniformity Normalized (GLRLM), Gray Level Non Uniformity Normalized LLL (GLRLM), Sum Entropy, Joint Energy, Entropy, Gray Level Non Uniformity Normalized (GLDM), Joint Energy, Gray Level Non Uniformity Normalized LLL (GLDM), Uniformity LLL, Sum Entropy LLL, and Uniformity, Zone Entropy HHH, Size Zone Non Uniformity Normalized HHH, and Small Area Emphasis HHH, Strength, Coarseness HLL, Coarseness, Coarseness LHL, Coarseness LLL, Coarseness LLH, Coarseness HHH, Coarseness HLH, Coarseness HHL, and Coarseness LHH, Cluster Tendency LLL, Cluster Tendency, Sum of Squares LLL, Mean Absolute Deviation LLL, Gray Level Variance LLL (GLDM), Variance LLL, Gray Level Variance LLL (GLRLM), Gray Level Variance (GLRLM), Robust Mean Absolute Deviation LLL, Gray Level Variance (GLDM), Variance, Mean Absolute Deviation, Cluster Prominence, Sum Entropy LLL, Interquartile Range LLL, Gray Level Variance LLL (GLSZM), Sum of Squares, Robust Mean Absolute Deviation, Sum Entropy, Interquartile Range, Cluster Prominence LLL, Entropy LLL, 10th Percentile LLL, 10th Percentile, Size Zone Non Uniformity LLL, Dependence Non Uniformity HLL, Gray Level Non Uniformity HLL (GLSZM), Gray Level Non Uniformity (GLSZM), Run Length Non Uniformity HHL, Run Length Non Uniformity LHL, Dependence Non Uniformity LHL, Dependence Non Uniformity, Run Length Non Uniformity HLH, Busyness, Run Length Non Uniformity LLH, Dependence Non Uniformity LLH, Dependence Non Uniformity LLL, Size Zone Non Uniformity, Energy HLL, Run Length Non Uniformity LHH, Size Zone Non Uniformity HLL, Gray Level Non Uniformity LLH (GLSZM), Gray Level Non Uniformity LHL (GLSZM), Gray Level Non Uniformity LLL (GLSZM), Run Length Non Uniformity HLL, Gray Level Non Uniformity HLH (GLSZM), Gray Level Non Uniformity HHL (GLSZM), Run Length Non Uniformity, and Run Length Non Uniformity HHH. The radiomic features may be selected from the radiomic features of clusters 1 to 9, wherein: cluster 1 consists of Short Run High Gray Level Emphasis, High Gray Level Emphasis, High Gray Level Run Emphasis, Autocorrelation, Sum Average, Joint Average, and High Gray Level Zone Emphasis; cluster 2 consists of Skewness, Skewness LLL, Kurtosis, 90th Percentile, 90th Percentile LLL, Median LLL, Kurtosis LLL, and Median; cluster 3 consists of Run Entropy, Dependence Entropy LLL, Dependence Entropy, Zone Entropy LLL, Run Entropy LLL, and Mean LLL; cluster 4 consists of Small Area Low Gray Level Emphasis, Low Gray Level Zone Emphasis, Short Run Low Gray Level Emphasis, Low Gray Level Run Emphasis, Low Gray Level Emphasis, Small Dependence Low Gray Level Emphasis, Gray Level Variance LLL (GLSZM), Gray Level Variance (GLDM), Variance, Gray Level Variance (GLDM), Difference Variance LLL, Gray Level Variance LLL (GLRLM), Variance LLL, Gray Level Variance LLL (GLDM), Sum of _{Squares, Contrast LLL, Mean Absolute Deviation, Interquartile Range, Robust} Mean Absolute Deviation, Long Run Low Gray Level Emphasis, Difference Variance, Gray Level Variance (GLSZM), Inverse Difference Moment Normalized, Mean Absolute Deviation LLL, Sum of Squares LLL, and Contrast; cluster 5 consists of Zone Entropy, Gray Level Non Uniformity Normalized (GLRLM), Gray Level Non Uniformity Normalized LLL (GLRLM), Sum Entropy, Joint Energy, Entropy, Gray Level Non Uniformity Normalized (GLDM), Joint Energy, Gray Level Non Uniformity Normalized LLL (GLDM), Uniformity LLL, Sum Entropy LLL, and Uniformity; cluster 6 consists of Zone Entropy HHH, Size Zone Non Uniformity Normalized HHH, and Small Area Emphasis HHH; cluster 7 consists of Strength, Coarseness HLL, Coarseness, Coarseness LHL, Coarseness LLL, Coarseness LLH, Coarseness HHH, Coarseness HLH, Coarseness HHL, and Coarseness LHH; cluster 8 consists of Cluster Tendency LLL, Cluster Tendency, Sum of Squares LLL, Mean Absolute Deviation LLL, Gray Level Variance LLL (GLDM), Variance LLL, Gray Level Variance LLL (GLRLM), Gray Level Variance (GLRLM), Robust Mean Absolute Deviation LLL, Gray Level Variance (GLDM), Variance, Mean Absolute Deviation, Cluster Prominence, Sum Entropy LLL, Interquartile _{Range LLL, Gray Level Variance LLL (GLSZM), Sum of Squares, Robust Mean} Absolute Deviation, Sum Entropy, Interquartile Range, Cluster Prominence LLL, Entropy LLL, 10th Percentile LLL, 10th Percentile; and cluster 9 consists of Size Zone Non Uniformity LLL, Dependence Non Uniformity HLL, Gray Level Non Uniformity HLL (GLSZM), Gray Level Non Uniformity (GLSZM), Run Length Non Uniformity HHL, Run Length Non Uniformity LHL, Dependence Non Uniformity LHL, Dependence Non Uniformity, Run Length Non Uniformity HLH, Busyness, Run Length Non Uniformity LLH, Dependence Non Uniformity LLH, Dependence Non Uniformity LLL, Size Zone Non Uniformity, Energy HLL, Run Length Non Uniformity LHH, Size Zone Non Uniformity HLL, Gray Level Non Uniformity LLH (GLSZM), Gray Level Non Uniformity LHL (GLSZM), Gray Level Non Uniformity LLL (GLSZM), Run Length Non Uniformity HLL, Gray Level Non Uniformity HLH (GLSZM), Gray Level Non Uniformity HHL (GLSZM), Run Length Non Uniformity, and Run Length Non Uniformity HHH. Preferably at least two radiomic features are selected; preferably wherein the at least two radiomic features are each selected from different clusters. The definitions of the radiomic features referred to herein are generally well understood within the field of radiomics by reference to their name only. However, for ease or reference definitions of the features used herein are provided in Tables R1 to R7 below. The radiomic features in Tables R1 to R7 are defined in accordance with the radiomic features used by the Pyradiomics package (http://pyradiomics.readthedocs.io/en/latest/features.html, see van Griethuysen, J. J. M., Fedorov, A., Parmar, C., Hosny, A., Aucoin, N., Narayan, V., Beets-Tan, R. G. H., Fillon-Robin, J. C., Pieper, S., Aerts, H. J. W. L. (2017). Computational _{Radiomics System to Decode the Radiographic Phenotype. Cancer Research,} 77(21), e104–e107. https://doi.org/10.1158/0008-5472.CAN-17-0339). Most features defined in Tables R1 to R7 are in compliance with feature definitions as described by the Imaging Biomarker Standardization Initiative (IBSI), which are available in Zwanenburg et al. (2016) (Zwanenburg, A., Leger, S., Vallières, M., _{and Löck, S. (2016). Image biomarker standardisation initiative - feature} definitions. In eprint arXiv:1612.07003 [cs.CV]). Where a definition provided below does not comply exactly from the IBSI definition, it should be understood that either definition could be used in accordance with the invention. Ultimately, the precise mathematical definition of the radiomic features is not crucial because slight modifications do not affect the general properties of the image that are measured by each of the features. Thus, slight modifications to the features (for example, the addition or subtraction of constants or scaling) and alternative definitions of the features are intended to be encompassed by the present invention. a. First Order Statistics These statistics describe the central tendency, variability, uniformity, asymmetry, skewness and magnitude of the attenuation values in a given region of interest (ROI), disregarding the spatial relationship of the individual voxels. As such, they describe quantitative and qualitative features of the whole ROI (PVR). A total of 19 features were calculated for each one of the eight wavelet transformations and the original CT image, as follows: Let: ^_{X be the attenuation or radiodensity values (e.g. in HU) of a set} _{of Np voxels included in the region of interest (ROI)} _{^ P(i) be the first order histogram with Ng discrete intensity levels,} _{where Ng is the number of non-zero bins, equally spaced from 0 with a} width. ^_{p(i) be the normalized first order histogram and equal to} P(^) ^_^ _{^ c is a value that shifts the intensities to prevent negative values in X. This} ensures that voxels with the lowest gray values contribute the least to Energy, instead of voxels with gray level intensity closest to 0. Since the HU range of adipose tissue (AT) within the PVR (-190 to -30 HU) does not include zero, c may be set at c=0. Therefore, higher energy corresponds to less radiodense AT, and therefore a higher lipophilic content. ^_{^ is an arbitrarily small positive number (e.g. ≈ 2.2×10−16)} Table R1: First-order radiomic features for PVR characterization _{Radiomic feature Interpretation} ^_{^ Energy is a measure of the magnitude of voxel} _{^nergy = ^(X(^) + ^)^} values in an image. A larger value implies a ^^^ greater sum of the squares of these values. ^_{^ Total Energy is the value of Energy feature} _{Total Energy = ^^^^^^ ^(X(^) + ^)}^ scaled by the volume of the voxel in cubic mm. ^^^ ^_{^ Entropy specifies the uncertainty/randomness} _{Entropy = − ^ ^(^)log^ (^(^) + ^)} in the image values. It measures the average ^^^ amount of information required to encode the image values M_{inimum = ^^^(X) The minimum gray level intensity within the} ROI. _{The 10th percentile of X The 10th percentile of X} _{The 90th percentile of X The 90th percentile of X} _{Maximum = ^^^(X) The maximum gray level intensity within the} ROI. ^{^} _^ 1 The average (mean) gray level intensity within M_{ean =} ^ _{^ X(^)} _^ the ROI. ^^^ _{Median The median gray level intensity within the ROI.} _{Interquartile range = P^^ − P^^ Here P25 and P75 are the 25} ^th and 75^th percentile of the image array, respectively. R_{ange = ^^^(X) − ^^^(X) The range of gray values in the ROI.} ^_^ 1 _{Mean Absolute Deviation (MAD) is the mean} _{MAD =} ^ _{^ |X(^) − ^|} _^ distance of all intensity values from the Mean ^^^ Value of the image array. ^^^^^^ 1 _{Robust Mean Absolute Deviation (rMAD) is} _{rMAD =} ^ _{^ |X^^^^^(^) − ^^^^^^|} _^^^^^ the mean distance of all intensity values from ^^^ the Mean Value calculated on the subset of image array with gray levels in between, or equal to the 10^th and 90^th percentile. ^_{^ Root Mean Squared (RMS) is the square-root} _{RMS = ^} 1 ^_{(X(^) + ^ ^} of the mean of all the squared intensity values. ^ ₎ _^ ^^^ It is another measure of the magnitude of the image values. This feature is volume- confounded, a larger value of c increases the effect of volume-confounding. S_{kewness =} ^_{^ Skewness measures the asymmetry of the} ^^{^} 1 ^{^} _^ distribution of values about the Mean value. ^ _{^ (X(^) − ^)^} Depending on where the tail is elongated a = _^ nd ^^^ the mass of the distribution is concentrated, (^ 1 ^{^} _^ _^ ^ ^{^} this value can be positive or negative. (Where ^_{^ (X(^) − ^)} ) ^^^ μ₃ is the 3^rd central moment). 1 _^ ^{^} _^ _{(X(^) − ^ ^ Kurtosis is a measure of the ‘peakedness’ of} _{Kurtosis =} ^_^ ^ = _{^ )} ^^^ the distribution of values in the image ROI. A ^^{^} ( 1 ^{^} _^ higher kurtosis implies that the ma ^_{^ ^ (X(^) − ^} )^{^})^{^} ss of the ^^^ distribution is concentrated towards the tail(s) rather than towards the mean. A lower kurtosis implies the reverse: that the mass of the distribution is concentrated towards a spike near the Mean value. (Where μ4 is the 4th central moment). ^{^} _{^ Variance is the mean of the squared distanc} _{Variance =} 1 _es ^ _{^(X(^) − ^)^} _^ of each intensity value from the Mean value. ^^^ This is a measure of the spread of the distribution about the mean. ^_{^ Uniformity is a measure of the sum of the} _{Uniformity = ^ ^(^)^} squares of each intensity value. This is a ^^^ measure of the heterogeneity of the image array, where a greater uniformity implies a greater heterogeneity or a greater range of discrete intensity values. b. Shape-related Statistics Shape-related statistics describe the size and shape of a given ROI, without taking into account the attenuation values of its voxels. Since they are independent of the gray level intensities, shape-related statistics were consistent across all wavelet transformation and the original CT image, and therefore were only calculated once. These were defined as follows: Let: _{V be the volume of the ROI in mm3} _{A be the surface area of the ROI in mm2} Table R2: Shape-related radiomic features for PVR characterization R_{adiomic feature Interpretation} ^ _{The volume of the ROI V is approximated by} _{^^^^^^ = ^ ^^} multiplying the number of voxels in the ROI by ^^^ the volume of a single voxel V_i. ^ _{Surface Area is an appro} _{^^^^^^^ ^^^^ = ^}1 _{ximation of the} surface of the ROI in mm², calculated using a 2_{|a^b^ × a^c^|} ^^^ _{marching cubes algorithm, where N is the} number of triangles forming the surface mesh of the v to hlume (ROI), a_ib_i and a_ic_i are the edges o_{f the i triangle formed by points ai, bi and ci.} _{Surface to volume ratio =} ^ Here, a lower value indicates a more compact ^ (sphere-like) shape. This feature is not dimensionless, and is therefore (partly) dependent on the volume of the ROI. ^ S_{phericity =} √36^^^{^} Sphericity is a measure of the roundness of ^ the shape of the tumor region relative to a sphere. It is a dimensionless measure, independent of scale and orientation. The value range is 0<sphericity≤1, where a value of 1 indicates a perfect sphere (a sphere has the smallest possible surface area for a given volume, compared to other solids). _{Volume Number Total number of discrete volumes in the ROI.} _{Voxel Number Total number of discrete voxels in the ROI.} _{Maximum 3D diameter Maximum 3D diameter is defined as the largest} pairwise Euclidean distance between surface voxels in the ROI (Feret . _{Maximum 2D diameter (Slice) Maximum 2D diameter (Slice) is defined as the} largest pairwise Euclidean distance between ROI surface voxels in the row-column (generally the axial) plane. _{Maximum 2D diameter (Column) Maximum 2D diameter (Column) is defined as} the largest pairwise Euclidean distance between ROI surface voxels in the row-slice (usually the coronal) plane. _{Maximum 2D diameter (Row) Maximum 2D diameter (Row) is defined as the} largest pairwise Euclidean distance between tumor surface voxels in the column-slice (usually the sagittal) plane. M_{ajor axis = 4^^major} λmajor is the length of the largest principal component axis M_{inor axis = 4^^minor λminor is the length of the second largest} principal component axis L_{east axis = 4^^least λleast is the length of the smallest principal} component axis ^_min Here, λ_major and λ_minor are the lengths of the E_{longation = ^ or} ^_{major largest and second largest principal} component axes. The values range between 1 (circle-like (non-elongated)) and 0 (single point or 1 dimensional line). Here, λmajor and λmino are the lengths of the t_{ness = ^} ^_lea r F_{la st} ^_{major largest and smallest principal component axes.} The values range between 1 (non-flat, sphere- like) and 0 (a flat object). c. Gray Level Co-occurrence Matrix (GLCM) In simple words, a GLCM describes the number of times a voxel of a given _{attenuation value i is located next to a voxel of j. A GLCM of size Ng×Ng describes} the second-order joint probability function of an image region constrained by the mask and is defined as P(i,j|δ,θ). The (i,j)^th element of this matrix represents the _{number of times the combination of levels i and j occur in two pixels in the image,} _{that are separated by a distance of δ pixels along angle θ. The distance δ from} the center voxel is defined as the distance according to the infinity norm. For δ=1, this results in 2 neighbors for each of 13 angles in 3D (26-connectivity) and for _{δ=2 a 98-connectivity (49 unique angles). In order to get rotationally invariant} results, statistics are calculated in all directions and then averaged, to ensure a symmetrical GLCM. Let: _{^ be an arbitrarily small positive number (e.g. ≈2.2×10−16)} _{P(i,j) be the co-occurrence matrix for an arbitrary δ and θ} _{p(i,j) be the normalized co-occurrence matrix and equal to} P(^,^) ∑P(^,^) Ng be the number of discrete intensity levels in the image _{^ ^} ^_^ _{^(^) =} ^^^ _{^(^, ^)} be the marginal row probabilities _{^^(^) = ^} ^_^ ^^^ _{^(^, ^)} be the marginal column probabilities _{μ be the mean gray level intensit ^} ^_^ _{x y of px and defined as ^^ =} ^^^ _^^(^)^ ^_^ _{μy be the mean gray level intensity of py and defined as ^^ = ^ ^^(^)^} ^^^ σx be the standard deviation of px σy be the standard deviation of py ^_^ ^_^ _{^^^^(^) = ^ ^ ^(^, ^) , where ^ + ^ = ^, and ^ = 2,3, … ,2^^} ^^^ ^^^ ^_^ ^_^ _{^^^^(^) = ^ ^ ^(^, ^) , where |^ − ^| = ^, and ^ = 0,1, … , ^^ − 1} ^^^ ^^^ _{^^ = − ^} ^_^ ^^^ _{^^(^)log^ (^^(^) + ^)} be the entropy of px _{^^ = − ^} ^_^ ^^^ _{^^(^)log^ (^^(^) + ^)} be the entropy of py ^_^ ^_^ _{^^^1 = − ^ ^ ^(^, ^)log^ (^^(^)^^(^) + ^)} ^^^ ^^^ ^_^ ^_^ _{^^^2 = − ^ ^ ^^(^)^^(^)log^ (^^(^)^^(^) + ^)} ^^^ ^^^ For distance weighting, GLCM matrices are weighted by weighting factor W and then summed and normalised. Weighting factor W is calculated for the distance _{between neighbouring voxels by ^ = ^^‖^‖}^ , where d is the distance for the associated angle. Table R3: Gray Level Co-occurrence Matrix (GLCM) statistics for PVR characterization R_{adiomic feature Interpretation} ^_{^ Autocorrelation is a} ^_^ measure of the magnitude of A_{utocorrelation = ^ ^(^, ^)^^} the fineness and coarseness ^^^ of texture. ^^^ ^_^ Returns the mean gray level ^_^ intensity of the i distribution. J_{oint average = ^^ = ^ ^(^, ^)^} ^^^ ^^^ Cluster prominence _{Cluster Prominence is a} ^_^ measure of the skewness ^_^ and asymmetry of the =_{^(^ + ^ − ^ ^} _{^ − ^^) ^(^, ^)} GLCM. A higher value implies more asymmetry ^^^ around the mean while a ^^^ lower value indicates a peak near the mean value and less variation around the mean. Cluster tendency _{Cluster Tendency is a} ^_^ measure of groupings of ^_^ voxels with similar gray-level =_{^(^ + ^ − ^^ − ^^)^^(^, ^)} values. ^^^ ^^^ ^_{^ Cluster Shade is a measure} ^_^ of the skewness and C_{luster shade = ^(^ + ^ − ^ ^} _{^ − ^^) ^(^, ^)} uniformity of the GLCM. A ^^^ higher cluster shade implies greater asymmetry about the ^^^ mean. ^_{^ Contrast is a measure of the} ^_^ local intensity variation, C_{− ^)^^(^, ^)} favoring values away from ^^^ the diagonal (i=j). A larger value correlates with a ^^^ greater disparity in intensity values among neighboring voxels. ^ ^ ∑ _{^ ^ ^ ^(^, ^)^^ − ^ ^ Correlation is a value} _{Correlation =} ^^^ ^^^ _{^ ^} between 0 (uncorrelated) ^_^(^)^_^(^) and 1 (perfectly correlated) showing the linear dependency of gray level values to their respective voxels in the GLCM. ^_^^^ Difference Average D_{ifference average = ^ ^^^^^(^)} measures the relationship ^^^ between occurrences of pairs with similar intensity values and occurrences of pairs with differing intensity values. Difference entropy _{Difference Entropy is a} ^_^^^ measure of the =_{^ ^^^^(^)log^ (^^^^(^) + ^)} randomness/variability in ^^^ neighborhood intensity value differences. ^_^^^ _{Difference Variance is a} _{Difference variance = ^ (^ − ^^)^^^^^(^)} measure of heterogeneity ^^^ that places higher weights on differing intensity level pairs that deviate more from the mean. ^_{^ Joint energy is a measure of} ^_^ homogeneous patterns in J_{oint energy = ^(^(^, ^))^} the image. A greater joint ^^^ energy implies that there are more instances of intensity ^^^ value pairs in the image that neighbor each other at higher frequencies. (also known as Angular Second Moment). ^_{^ Joint entropy is a measure} ^_^ of the randomness/variability J_{oint entropy = − ^ ^(^, ^)log^ (^(^, ^) + ^)} in neighborhood intensity ^^^ values. ^^^ ^_{^^ − ^^^1} _{IMC 1 =} Informational measure of ^_{^^{^^, ^^}} correlation 1 I_{MC 2 = ^1 − ^^^(^^^^^^^^)} _{measure of} correlation 2 ^_^ ^_^ IDM (inverse difference ^_{(^, ^) moment a.k.a Homogeneity} _{^ ^} 2) is a measure of the local 1_{+ |^ − ^|} homogeneity of an image. ^^^ IDM weights are the inverse ^^^ of the Contrast weights (decreasing exponentially from the diagonal i=j in the GLCM). ^_^ ^_^ IDMN difference is a I_{DMN =} local |_{^ − ^|^} +₍ ^_^ ^{^} ) IDMN weights are the ^^^ ^^^ inverse of the Contrast weights (decreasing exponentially from the diagonal i=j in the GLCM). Unlike 2, IDMN square of the between values the square of of discrete _{ID (inverse difference a.k.a.} Homogeneity 1) is another measure of the local homogeneity of an image. With more uniform gray levels, the denominator will remain low, resulting in a higher overall value. IDN (inverse difference normalized) is another = measure of the local 1_{+ (|^ − ^|} homogeneity of an image. ^_^ ) Unlike Homogeneity 1, IDN ^^^ ^^^ normalizes the difference between the neighboring intensity values by dividing over the total number of discrete intensity values. ^_^ ^_^ _{^(^, ^)} _{Inverse variance = ^} _{|^ − ^|^ , ^ ≠ ^} ^^^ ^^^ M_{aximum probability = ^^^(^(^, ^)) Maximum Probability is} occurrences of the most predominant pair of neighboring intensity values (also known as Joint maximum). ^^_^ Sum Average measures the S_{um average = ^ ^^^^(^)^} relationship between ^^^ occurrences of pairs with lower intensity values and occurrences of pairs with higher intensity values. ^^_{^ Sum Entropy is a sum of} _{Sum entropy = ^ ^^^^(^)log^ (^^^^(^) + ^)} neighborhood intensity value ^^^ differences. ^_{^ Sum of Squares or Variance} ^_^ is a measure in the S_{um squares = ^(^ − ^ ^} _{^) ^(^, ^)} distribution of neighboring ^^^ intensity level pairs about the mean intensity level in the ^^^ GLCM. (Defined by IBSI as Joint Variance). d. Gray Level Size Zone Matrix (GLSZM) A Gray Level Size Zone (GLSZM) describes gray level zones in a ROI, which are defined as the number of connected voxels that share the same gray level intensity. A voxel is considered connected if the distance is 1 according to the infinity norm (26-connected region in a 3D, 8-connected region in 2D). In a gray _{level size zone matrix P(i,j) the (i,j)th element equals the number of zones with} _{gray level i and size j appear in image. Contrary to GLCM and GLRLM, the} GLSZM is rotation independent, with only one matrix calculated for all directions in the ROI. Let: Ng be the number of discreet intensity values in the image Ns be the number of discreet zone sizes in the image N_p be the number of voxels in the image ^ ^ N e number of zones in the ROI, which is equal to _^ _z be th ∑ ^^^ _{^ ^} ^^^ _{P(^, ^)} and 1 ≤ Nz ≤ Np _{P(i,j) be the size zone matrix} _{p(i,j) be the normalized size zone matrix, defined as ^(^, ^) =} P(^,^) ^_^ _{^ is an arbitrarily small positive number (e.g. ≈2.2×10−16).} Table R4: Gray Level Size Zone Matrix (GLSZM) statistics for PVR characterization _{Radiomic feature Interpretation} ^_^ ^ _{SAE (small area emphasis) is a} ∑ _{^ P(^, ^)} ^^^ ^ measure of the distribution of small size ^^{^} ^ zones, with a greater value indicative of S_{AE =} ^^ ^_^ smaller size zones and more fine textures. ^ ∑ _^ ^_{^ ^ LAE (large area em} ^^^ _{^ P(^, ^)^ phasis) is a} _{LAE =} ^^^ measure of the distribution of large area ^_^ size zones, with a greater value indicative of larger size zones and more coarse textures. ^ ∑ _^ ^_^ GLN (gray level non-uniformi G_{LN =} ^^^ _(^ ^^^ _{P(^, ^)} )^{^} ty) measures the variability of gray-level ^_^ intensity values in the image, with a lower value indicating more homogeneity in intensity values. ^ ∑ _^ ^_^ GLNN (gray level non-unif ^^^ _(^ ^^ _{P(^, ^)} )^{^} ormity G_{LNN =} ^ _{normalized) measures the variability of} ^_^ ^{^} gray-level intensity values in the image, with a lower value indicating a greater similarity in intensity values. This is the normalized version of the GLN formula. ∑^{^^} ^_^ SZN (size zone non-uniformity S_{ZN = ^^^ (^} ^^^ _{P(^, ^)} )^{^} ) measures the variability of size zone ^_^ volumes in the image, with a lower value indicating more homogeneity in size zone volumes. ∑^{^^} _(^ ^_{^ P(^, ^} ^{^} SZNN (size zone non-uniformity S_{ZNN = ^^^} ^^^ ₎ ) n_{ormalized) measures the variability of} ^_^ ^{^} size zone volumes throughout the image, with a lower value indicating more homogeneity among zone size volumes in the image. This is the normalized version of the SZN formula. Z_{one Percentage =} ^_{^ ZP (Zone Percentage) measures the} ^_^ coarseness of the texture by taking the ratio of number of zones and number of voxels in the ROI. Values are in range ^ ^_{^ ≤ ^^ ≤ 1, with higher values} indicating a larger portion of the ROI consists of small zones (indicates a more fine texture). ^_^ _{GLV = ^ ^} ^_{^ ^(^, ^)(^ − ^)^ Gray level variance (GLV) measures} , ^^^ ^^^ the variance in gray level intensities for _{where ^ = ^} ^{^} _^ _^ ^{^} _{^ ^(^, ^)^} the zones. ^^^ ^^^ ^_^ _{ZV = ^ ^} ^_{^ ^(^, ^)(^ − ^)^ Zone Variance (ZV) measures the} , where ^^^ ^^^ variance in zone size volumes for the ^ _{^ = ^} ^{^} ^ ^{^} _{^ ^(^, ^} zones. ^^ _)^ ^^^ ^ ^_{^ Zone Entropy (ZE) measures the} ^_^ uncertainty/randomness in the Z_{E = − ^ ^(^, ^)log^ (^(^, ^)} distribution of zone sizes and gray ^^^ levels. A higher value indicates more heterogeneneity in the texture patterns. ^^^ +_^) ^ ^_{^ P(^, ^} LGLZE (low gray level zone ∑ _^ ^ ₎ ^^^ ^^{^} _{emphasis) measures the distribution of} _{LGLZE =} ^^^ ^ lower gray-level size zones, with a ^ higher value indicating a greater proportion of lower gray-level values and size zones in the image. ^ ∑ _^ ^_{^ ^} HGLZE (high gray le H_{GLZE =} ^ _{^ P(^, ^)^} vel zone ^^ ^^^ _{emphasis) measures the distribution of} ^_^ the higher gray-level values, with a higher value indicating a greater proportion of higher gray-level values and size zones in the image. ^_^ ^ SALGLE (small area low gray l ∑ _{^ P(^, ^)} evel ^^^ ^ ^^{^}^^{^} _{emphasis) measures the proportion in} ^ the image of the joint distribution of S_{ALGLE =} ^^ ^_^ smaller size zones with lower gray-level values. ^_^ ^ ∑ _{^ P(^, ^)^^} SAHGLE (small area high gray level ^^^ ^ ^^{^} _{emphasis) measures the proportion in} the image of the joint distribution of S_{AHGLE =} ^^^ ^_^ smaller size zones with higher gray-level values. ^_^ ^ _{P(^, ^} LALGLE (low area low gray level ∑ _{^ ^)^} ^^^ ^ emphasis) measures the proportio ^^{^} n in the image of the joint distribution of L_{ALGLE =} ^^^ ^_^ larger size zones with lower gray-level values. ^ ∑ _^ ^_{^ ^ ^} _{LAHGLE =} ^^^ _^ ^^^ _{P(^, ^)^ ^} LAHGLE (low area high gray level emphasis) measures the proportion in ^_^ the image of the joint distribution of larger size zones with higher gray-level values. e. Gray Level Run Length Matrix (GLRLM) A Gray Level Run Length Matrix (GLRLM) describes gray level runs, which are defined as the length in number of pixels, of consecutive pixels that have the same gray level value. In a gray level run length matrix P(i,j|θ), the (i,j)^th element _{describes the number of runs with gray level i and length j occur in the image} (ROI) along angle θ. Let: Ng be the number of discreet intensity values in the image Nr be the number of discreet run lengths in the image Np be the number of voxels in the image _{Nz(θ) be the number of runs in the image along angle θ, which is equal to} ^ ∑ _^ ^^^ _^ ^_^ ^^^ _{P(^, ^|^)} and 1≤N_z(θ)≤N_p _{P(i,j|θ) be the run length matrix for an arbitrary direction θ} _{p(i,j|θ) be the normalized run length matrix, defined as ^(^, ^|^) =} P(^,^|^) ^_^(^) _{^ is an arbitrarily small positive number (e.g. ≈2.2×10−16)} By default, the value of a feature is calculated on the GLRLM for each angle separately, after which the mean of these values is returned. If distance weighting is enabled, GLRLMs are weighted by the distance between neighbouring voxels and then summed and normalised. Features are then the resultant matrix. The distance between neighbouring voxels is for each angle using the norm specified in ‘weightingNorm’ Table R5: Gray Level Run Length Matrix (GLRLM) statistics for PVR characterization R_{adiomic feature Interpretation} ^_^ ^ ∑ _{^ P(^, ^|^) SRE (Short Run Emphasis) is} ^^^ ^ a measure of the distribution of ^^{^} ^^ short run lengths, with a S_{RE =} ^ ^_^(^) greater value indicative of shorter run lengths and more fine textural textures. ^ ∑ _{^ ^} ^_{^ P(^, ^|^)^^} LRE (Long Run Emphasis) is L_{RE =} ^^^ ^^^ a measure of the distribution of ^_^(^) long run lengths, with a greater value indicative of longer run lengths and more coarse structural textures. ^ ∑ _^ ^_^ ^{^} GLN (Gray Level Non- G_{LN =} ^^^ _(^ ^^^ _{P(^, ^|^)} ) u_{niformity) measures the} ^_^(^) similarity of gray-level intensity values in the image, where a lower GLN value correlates with a greater similarity in intensity values. ^ ∑ _{^ (^} ^_{^ P(^,} ^{^} GLNN (Gray Level Non- G_{LNN =} ^^^ ^^^ _^|^) ) ^ uniformity Normalized) ^_^(^) measures the similarity of gray- level intensity values in the image, where a lower GLNN value correlates with a greater similarity in intensity values. This is the normalized version of the GLN formula. ∑_^^ ^_^ ^{^} RLN (Run Length Non- _{RLN = ^^^ (^} ^^^ _{P(^, ^|^)} ) uniformity) measures the ^_^(^) similarity of run lengths throughout the image, with a lower value indicating more homogeneity among run lengths in the image. ∑_{^^ (} ^_^ ^{^} RLNN (Run Length Non-_{LNN = ^^^ ^} ^^^ _{P(^, ^|^)} ) u_{niformity) measures the} ^_^(^)^{^} similarity of run lengths throughout the image, with a lower value indicating more homogeneity among run lengths in the image. This is the normalized version of the RLN formula. R_{P =} ^_^(^) RP (Run Percentage) ^_^ measures the coarseness of the texture by taking the ratio of number of runs and number of voxels in the ROI. Values are in ^ range ^_{^ ≤ ^^ ≤ 1, with higher} values indicating a larger portion of the ROI consists of short runs (indicates a more fine texture). ^_^ _^ ^_{^ ^(^, ^|^)(^ − ^)^} GLV (Gray Level Variance) , ^^^ ^^^ measures the variance in gray level intensity for the runs. ^_^ _{= ^ ^} ^_^ ^ _{^(^, ^|^)^} ^^^ ^^ ^_^ _^ ^_{^ ^(^, ^|^)( ^} RV (Run Variance) is a ^^^ ^^^ _{^ − ^)} , measure of the variance in runs for the run lengths. ^_^ _{= ^ ^} ^_^ ^^ _{^(^, ^|^)^} ^^^ ^ ^_{^ RE (Run Entropy) measures} ^_^ the uncertainty/randomness in _{RE = − ^ ^(^, ^|^)log^ (^(^, ^|^) + ^)} the distribution of run lengths ^^^ and gray levels. A higher value indicates more heterogeneity in ^^^ the texture patterns. ^ ^_{^ P(^,} LGLRE (low gray level run ∑ _^ ^^^ ^ _^|^) ^^^ ^^{^} _{emphasis) measures the} _{LGLRE =} distribution of low gray-level ^_^(^) values, with a higher value indicating a greater concentration of low gray-level values in the image. ^ ∑ _^ ^_{^ ^} HGLRE (high gray le H_{GLRE =} ^ _{^ P(^, ^|^)^} vel run ^^ ^^^ emphasis) measures the ^_^(^) distribution of the higher gray- level values, with a higher value indicating a greater concentration of high gray-level values in the image. ^_^ ^ SRLG ∑ _{^ P(^, ^|^)} LE (short run low gray ^^^ ^ level emphasis) measures the ^^{^}^^{^} ^^^ joint distribution of shorter run S_{RLGLE =} ^_^(^) lengths with lower gray-level values. ^_^ ^ SRH ∑ _{^ P(^, ^|^)^^} GLE (short run high gray ^^^ ^ ^^{^} _{level emphasis) measures the} joint distribution of shorter G_{LE =} ^ run S_RH ^^ ^_^(^) lengths with higher gray-level values. ^_^ ^ ∑ _{^ P(^, ^|^)^^} LRLGLRE (long run low gray ^^^ ^ ^^{^} _{level emphasis) measures the} joint distribution of long r L_{RE =} ^^ un L_RLG ^ ^_^(^) lengths with lower gray-level values. ^_^ ^ ∑ _{^ ^ ^} LRHGLRE (long run ^^ _{^ P(^, ^|^)^ ^} high L_{RHGLRE =} ^ ^^^ gray level run emphasis) ^_^(^) measures the joint distribution of long run lengths with higher gray-level values. f. Neighbouring Gray Tone Difference Matrix (NGTDM) Features A Neighbouring Gray Tone Difference Matrix quantifies the difference between a gray value and the average gray value of its neighbours within distance δ. The _{sum of absolute differences for gray level i is stored in the matrix. Let X^^be a set} _{∈ X^^ be the gray level of a voxel at postion} _{(^^ , ^^ , ^^), then the average gray level of the neigbourhood is:} ^_{^ = ^(^^ , ^^, ^^)} ^ ^ ^ 1 = _{^ ^^^(^^ + ^^ , ^^ + ^^, ^ + ^ )} , ^ _{^ ^} ^_^^^^ ^_^^^^ ^_^^^^ w_{here (^^ , ^^, ^^) ≠ (0,0,0) and ^^^(^^ + ^^ , ^^ + ^^, ^^ + ^^) ∈ X^^} _{Here, W is the number of voxels in the neighbourhood that are also in Xgl.} Let: _{ni be the number of voxels in Xgl with gray level i} _{Nv,p be the total number of voxels in Xgl and equal to ∑^^ (i.e. the number of} _{voxels with a valid region; at least 1 neighbor). ^^,^ ≤ ^^, where Np is the total} number of voxels in the ROI. p_i be the gray level probability and equal to ^_^/^_^ ^_^ _{^^ = {^ ^^ − ^^^ for ^^ ≠ 0} be the sum of absolute differences for gray level 0_{for ^^ = 0} i N_g be the number of discreet gray levels Ng,p be the number of gray levels where pi≠0 T_{able R6: Neigbouring Gray Tone Difference Matrix (NGTDM) for PVR} characterization R_{adiomic feature Interpretation} _{^^^^^^^^^^ =} 1 _{Coarseness is a measure of} _∑ ^_^ average difference between the ^^^ _^^^^ center voxel and its neighbourhood and is an indication of the spatial rate of change. A higher value indicates a lower spatial change rate and a locally more uniform texture. ^^^^^^^^ _{Contrast is a measure of the} ^_^ spatial intensity change, but is also ^_^ _æ dependent on the overall gray level 1 = ^_{^ ^ ^ (^} dynamic range. Contrast is high ç _{^ ^} _{^,^^^^,^ − 1^} when both the dynamic range and ^^^ the spatial change rate are high, i.e. è ^^^ an image with a large range of gray ^_^ levels, with large changes between ^ ö 1 voxels and their neighbourhood. −_^) ÷ ^ ^ _{^ ^^ ^ ,} _^,^ ^^^ ø w_{here ^^ ≠ 0, ^^ ≠ 0} _∑ ^_^ A measure of the change from a ^_{^^^^^^^ =} ^^^ _^^^^ ^_^ , pixel to its neighbour. A high value ^ ∑ _^ for busyness indicates a ‘busy ^^ _{^ ^^^^ − ^^^^} ’ ^ ^^^ image, with rapid changes of intensity between pixels and its neighbourhood. w_{here ^^ ≠ 0, ^^ ≠ 0} ^^^^^^^^^^ An image is considered complex ^_^ ^_^ when there are many primitive ^_{^^ + ^ ^} components in the image, i.e. the 1 = _{^|^ − ^| ^ ^ ^} image is non-uniform and the ^_{^,^ ^} , re are ^_{+ ^^} many rapid changes in gray level ^^^ intensity. ^^^ w_{here ^^ ≠ 0, ^^ ≠ 0} ^^^^^^^^ _{Strength is a measure of the} ^_^ ^_^ _{^ (^ + ^ )(^ − ^)}^ primitives in an image. Its value is ∑ ^^^ _{^ ^} ^^^ high when the primitives are easily = ^ , defined and visible, i.e. an image ∑_^ ^^^ _^^ with slow change in intensity but more large coarse differences in gray level intensities. w_{here ^^ ≠ 0, ^^ ≠ 0} g. Gray Level Dependence Matrix (GLDM) A Gray Level Dependence Matrix (GLDM) quantifies gray level dependencies in an image. A gray level dependency is defined as the number of connected voxels _{within distance δ that are dependent on the center voxel. A neighbouring voxel} _{with gray level j is considered dependent on center voxel with gray level i if |i−j|≤α.} _{In a gray level dependence matrix P(i,j) the (i,j)th element describes the number} _{of times a voxel with gray level i with j dependent voxels in its neighbourhood} appears in image. N_g be the number of discreet intensity values in the image N_d be the number of discreet dependency sizes in the image N_z be the number of dependency zones in the image, which is equal to ^ ∑ _^ ^^^ _^ ^_^ ^^^ _{P(^, ^)} _{P(i,j) be the dependence matrix} _{p(i,j) be the normalized dependence matrix, defined as ^(^, ^) =} P(^,^) ^_^ Table R7: Gray Level Dependence Matrix (GLDM) statistics for PVR characterization R_{adiomic feature Interpretation} ^ ^_^ SDE (Small Dependence Emphasis): ∑ _{^ P(^, ^)} ^^^ ^ ^^ ^^{^} A measure of the distribution of small ^_{^^ =} ^ ^_^ dependencies, with a greater value indicative of smaller dependence and less homogeneous textures. ^ ∑ _^ ^_{^ ^} LDE (Large Dependence ^_{^^ =} ^^^ _^ ^^^ _{P(^, ^)^} Emphasis): A measure of the distribution of large ^_^ dependencies, with a greater value indicative of larger dependence and more homogeneous textures. ^ ∑ _^ ^_^ GLN (Gray Level Non-Uniformity ^_{^^ =} ^ _{(^ P(^, ^)} )^{^} ): ^^ ^^^ ^ ∑ _^ Measures the similarity of gray-level ^^{^} ^^^ ∑ _{^^^ P(^, ^)} intensity values in the image, where a lower GLN value correlates with a greater similarity in intensity values. : a lower ^ of dependence throughout the image, with a lower value indicating more homogeneity among dependencies in the image. This is the normalized version of the DLN formula. ^_^ GLV Variance): Measures ^_^ the variance in grey level in the image. ^_{^^ = ^ ^(^, ^)(^ − ^)^} , ^^^ ^^^ ^_^ ^_^ _{(^, ^)} ^^^ ^_^ DV (Dependence Variance): ^_^ Measures the variance in dependence ^_{^ = ^ ^(^, ^)(^ − ^)^} , size in the image. ^^^ ^^^ ^_^ ^_^ _{where ^ = ^ ^^(^, ^)} ^^^ ^^^ ^_^ DE (Dependence Entropy): Measures ^_^ the entropy in dependence size in the ^_{^ = − ^ ^(^, ^)log^ (^(^, ^)} image. ^^^ ^^^ +_^) ^ ^_{^ P(^} LGLE (Low Gray Level Emphasis): ∑ _{^ , ^)} ^^^ ^ ^^^ ^^{^} Measures the distribution of low gray- ^_{^^^ =} ^_^ level values, with a higher value indicating a greater concentration of low gray-level values in the image. ^ ∑ _^ ^_{^ ^} HGLE (High G ^_{^^^ =} ^^^ _^ ^^^ _{P(^, ^)^} ray Level Emphasis): Measures the distribution of the higher ^_^ gray-level values, with a higher value indicating a greater concentration of high gray-level values in the image. ^_^ ^ ∑ _{^ P(^, ^)} SDLGLE (Small Dependence Low ^^^ ^ Gray Level Emphasis): Measures the ^^{^}^^{^} ^^^ joint distribution of small dependence ^_{^^^^^ =} ^_^ with lower gray-level values. ^_^ ^ ∑ _{^ P(^, ^)^^} SDHGLE (Small Dependence High ^^^ ^ Gray Level Emphasis): Measures ^^{^} the ^^^ joint distribution of small dependence ^_{^^^^^ =} ^_^ with higher gray-level values. ^_^ ^ LD ∑ _{^ P(^, ^)^^} LGLE (Large Dependence Low ^^^ ^ Gray Level Emphasis): Measures th ^^{^} e joint distribution of large depend ^₌ ^^ ence ^_^^^^ ^ ^_^ with lower gray-level values. ^ ∑ _^ ^_{^ ^ ^} LDHGLE (Large Dependence High ^_{^^^^^ =} ^^^ _^ ^^^ _{P(^, ^)^ ^} Gray Level Emphasis): Measures the ^_^ joint distribution of large dependence with higher gray-level values. _{The term “vascular state” denotes the biological condition of a blood vessel or} _{segment thereof and the tissue structures associated with the vessel or segment,} for example, the perivascular tissue. This vascular status includes the degree and type of any abnormality present. An example vessel status would be inflammation, which is associated with increased lipolysis and inhibited adipogenesis in the perivascular tissue. An alternative vascular status would be fibrosis, which is associated with reduced lumen diameter and arterial wall thickening due to excessive deposition of extracellular matrix along with collagen deposition in the _{perivascular space. The number of vessels or segments thereof with “vascular} state” in step (c) in the context of the methods of the invention is the number or _{vessels or segments having “adverse vascular state” e.g. states which are} _{detrimental to the vascular and overall health of a subject.} As used herein, the term “perivascular” refers to the space that surrounds a blood _{vessel. The term “perivascular tissue” or “perivascular space” refers to the tissue} that surrounds a blood vessel, and may include perivascular adipose tissue (PVAT). The perivascular space is in the plane of the blood vessel circumference which extends to a distance beyond the outer wall of the vessel to the perivascular _{region. Methods of determining the perivascular space surrounding blood vessels} are not limited, and are well known to the skilled person in the art. For example, beyond the outer wall of the vessel may depend on the analysed, and its spatial positioning in relation to other bodily tissues, or other surrounding vessels. For instance, if the particular vessel being analysed is positioned close to another type of body tissue, the perivascular space should not include voxels corresponding to the other body The distance extending beyond the outer wall of the vessel may also be determined by the radio densities in the image (as for radial extent) or based on patient anatomy. For example, thinner perivascular tissue more distally along the _{vessel; or to prevent inclusion of other anatomical features e.g. to exclude tissue} outside the pericardial sack in the case of the coronary arteries. Preferably, the distance extending beyond the outer wall of the vessel is determined by the processing system (i.e. this step is computer implemented), which preferably calculates this distance based on reference data sets. Alternatively, the distance may be manually determined or adjusted by an operator of the processing system. In one embodiment, the distance extending beyond the outer wall of the vessel may be: i_{) a standard distance that is not equal to or related to the diameter} or radius of the underlying vessel; i_{i) a distance which is a derivative of, a multiple of, or equal to the} radius or diameter of the underlying vessel; or i_{ii) from 0.1mm to 3cm, preferably from 0.2mm to 2.5cm, more} preferably from 0.3mm to 2.25cm, more preferably from 0.4mm to 2cm. In one embodiment, a value for the radiodensity of the perivascular tissue surrounding each blood vessel or segment thereof may involve determining a radiodensity value of one or more particular types of tissue, for example, one or more of: adipose tissue, fibrous plaque, calcified plaque, or water. _{The radiodensity value in (i) preferably comprises determining the radiodensity of} preferably perivascular adipose tissue (PVAT).Step (b) of the method preferably comprises comparing the value of an inflammation metric derived from the radiodensity value to a pre-determined cut off value, or using the absolute value, to determine whether each blood vessel or segment thereof is inflamed. _{For example, the inflammation metrics Fat Attenuation Index (FAI) and/or} _{perivascular FAI (FAIPVAT) can be determined, which reflects the standardized,} _{weighted average attenuation of a perivascular region around the human coronary} _{arteries. These metrics have been found to be a sensitive and dynamic biomarker} _{of coronary inflammation, and have been identified as a strong and independent} _{predictors of adverse cardiac events.} _{In a particularly preferred embodiment, step (b) comprises determining fat} attenuation index (FAI), and/or fat attenuation index of perivascular adipose tissue (FAI_PVAT). FAI and FAIPVAT are based on the understanding that in the presence of vascular inflammation, the release of pro-inflammatory molecules from the diseased _{vascular wall inhibits differentiation and lipid accumulation in pre-adipocytes within} the perivascular tissue (PVT), resulting in smaller, less differentiated and lipid-free adipocyte cells. This is associated with a shift in the radiodensity (measured as _{attenuation values) of PVT in computed tomography (CT) imaging from more} negative (closer to -190) to less negative (closer to -30) Hounsfield Unit (HU) _{values, which may be captured by the FAI and FAIPVAT. Both the biological} meaning and clinical value of the FAI have been extensively validated. These biomarkers are described in detail in WO 2016/024128 A1 and WO 2018/078395 A1, the contents of which are herein incorporated by reference. In the context of the present invention, an “average” value is understood to mean a central or typical value, and it can be calculated from a sample of measured values using formulas that are widely known and appreciated in the art. Preferably, the average is calculated as the arithmetic mean of the sample of attenuation values, but it can also be calculated as the geometric mean, the harmonic mean, the median or the mode of a set of collected attenuation values. The average value may be calculated by reference to data collected from all _{voxels within a concentric tissue layer or by reference to a selected population of} _{voxels within the concentric tissue layer, for example water- or adipose tissue- or} _{fibrous tissue- containing voxels.} The blood vessels from which the CT scan data is gathered are preferably _{coronary arteries. There are typically three main coronary arteries – the right} coronary artery, left anterior descending artery and left circumflex artery. These are illustrated schematically in Figure 5. _{The computer implemented method of the invention may further comprise} _{determining the length of each blood vessel or segment thereof which has disease} and/or (adverse) vascular state, preferably selected from: inflammation, oedema, fibrosis, vascularity, lipolysis and/or adipogenesis, and using this value in step (c) _{of claim 1 – thus illustrating the total level of disease and/or (adverse) vascular} _{state in a given subject.} _{For instance, this step may comprise determining the total length of blood vessel} or segment thereof that has disease and/or vascular state, preferably selected from: inflammation, oedema, fibrosis, vascularity, lipolysis and/or adipogenesis, _{across each of the coronary arteries and/or each of the carotid arteries, as outlined} above. Preferably the total length of inflamed blood vessel. The severity of disease and/or vascular state, preferably selected from: inflammation, oedema, fibrosis, vascularity, lipolysis and/or adipogenesis, _{preferably inflammation, in each length of blood vessel or segment thereof may} also be quantified, and used in step (c). _{In a preferred embodiment, the method may further comprise determining the total} _{length of blood vessel or segment thereof that has disease and/or vascular state,} _{preferably selected from inflammation, oedema, fibrosis, vascularity, lipolysis} _{and/or adipogenesis, most preferably inflammation, across each of the coronary} arteries and/or each of the carotid arteries weighted by: (_{i) the lumen diameter in each portion of a vessel;} _{(ii) the lumen volume of each portion of the vessel;} _{(iii) the percentage of fractional myocardial mass; or} _{(iv) the total length of vessel distal to each portion of the vessel; and} using this value in step (c) of claim 1. The ‘portion’ of each vessel may relate to the numbered segment of the artery _{(e.g. as shown in Figure 5), or the thoracic or abdominal aortic segments as} described above. _{This embodiment of the invention enables the relative significance of each vessel} or segment thereof to be taken into account when determining the risk score. For example, the relative size and importance of each of the coronary arteries in determining a subject’s risk depends on not only the severity of disease and/or _{vascular state (e.g. inflammation) or number of vessels with disease and/or} _{vascular state (e.g. number of inflamed vessels), but also the significance of those} vessels or segments to the particular subject. The arrangement, size and lumen volume of each coronary artery widely varies _{between individuals. Some subjects, for example, have a very short RCA,} _{therefore a high level of disease and/or vascular state in this artery compared to} the other coronary arteries supplying the heart would be expected to not have as much of an impact on the patient’s risk of suffering a major cardiovascular event _{compared to the left coronary arteries having larger lumen diameter, lumen} volume, percentage of fractional myocardial mass, or total length of vessel distal to each portion of the vessel. The term “weighted by” means the weighted average, and is calculated by multiplying the weight (or probability) associated with a particular outcome with its _{associated quantitative outcome, and then summing all the products together. The} _{method may further comprise quantifying the plaque burden in each blood vessel} or in each vessel segment, and using this value in step (c) of claim 1. _{In a preferred embodiment, the method further comprises determining the total} _{number and/or length of blood vessel or segment thereof that has disease and/or} vascular state, preferably selected from inflammation, oedema, fibrosis, _{vascularity, lipolysis and/or adipogenesis, most preferably inflammation, across} each of the coronary arteries, each of the carotid arteries, and/or each of the thoracic and/or abdominal aortic branches or segments, weighted by the plaque burden in each portion of the vessel, and using this value in step (c) of claim 1. _{The plaque burden comprises determining:} _{(i) the total volume of plaque;} _{(ii) the volume of calcified plaque; and/or} _{(iii) the volume of non-calcified plaque; and/or.} To determine the output value indicating the disease and/or vascular state (e.g. MACE risk score), the value for the radiodensity and/or one or more radiomic features of the perivascular tissue surrounding each blood vessel or segment thereof is compared to a pre-determined threshold value. Alternatively, the absolute values are used. In one embodiment, the value for the radiodensity and/or one or more radiomic features of the perivascular tissue surrounding each blood vessel or segment thereof (e.g. the inflammation metric FAI Score) for a given subject is compared _{against the data of a patient population matched for age, gender and/or ethnicity.} If the blood vessels or segments thereof being analysed have a value for the radiodensity and/or one or more radiomic features above a particular percentile in relation to the population data, it is classed as having disease and/or (adverse) vascular state For example, a vessel or segment thereof may be classified as inflamed if it has a value for the radiodensity (e.g. FAI Score) in the 50th percentile or above, preferably the 60th percentile or above, preferable the 70th percentile or above, most preferably 75th percentile compared to the reference data set. The threshold value may be a distribution of values, more preferably the threshold value is vessel specific and/or specific to how proximal or distal along the vessel the value was derived. It is particularly preferred that the threshold value and/or output value is adjusted for biological factors, which may impact the expected values. Preferably, the output value is adjusted. The biological factors are preferably selected from one or more of the age of the subject, the gender of the subject, the ethnicity of the _{subject, the background adipocyte size, partial volume effects (resulting from the} _{interaction of the specific patient anatomy and performance of the imaging} system), and the type of blood vessel. _{The partial volume effects could be determined using an Expectation Maximisation} _{approach to iteratively refine the estimated “partial volume corrected” voxel} configuration which when convolved with the point spread function of the imaging system most closely matches the observed voxel configuration. Technical factors may also impact the expected radiodensity and/or radiomic values obtained. Accordingly, the pre-determined threshold value and/or the output value may be adjusted to take account of these technical factors, such as one or more of: the tube voltage of the CT scanner, reconstruction algorithms, _{scanner resolution, iodinated contrast agent, contrast type, injection rate, aortic} contrast opacification, left ventricular blood pool opacification, signal-to-noise, contrast-to-noise, milliamps, method of cardiac gating, single and multiple energy image acquisition, CT scanner type, heart rate, heart rhythm, or blood pressure. The corrections for biological and/or technical factors are preferably carried out by the processing system i.e. they are computer implemented. Such processing _{systems may recognise the need for, and implement, these corrections} automatically. In some embodiments, the inflammation metric (e.g. FAI score) may be generated using a trained machine learning or statistical model, wherein the input to the model comprises the value for the radiodensity and/or one or more radiomic features of the blood vessel or segments thereof (e.g. obtained in step (a) of the method). The model is typically a regression model such as a Cox proportional hazards regression model. However, other suitable machine learning or statistical models may be used as would be understood by the person skilled in the art. Preferably, the input to the model may further comprise at least one of: (_{i) one or more technical factors (e.g. of the CT scanner used to obtain} the radiodensity values); (ii) one or more biological factors of the patient and/or blood vessel. The one or more technical factors and the one or more biological factors may be as described above. Typically, the machine learning or statistical model may be a model as described in Oikonomou EK, Antonopoulos AS, Schottlander D, et al. Standardized measurement of coronary inflammation using cardiovascular computed tomography: integration in clinical care as a prognostic medical device. _{Cardiovasc Res 2021; 117(13): 2677-90.} The coefficients for each variable used in step (c) to generate the output value may be derived from Cox proportional hazard regression models. Alternatively, the cut-off points may be derived from received operating characteristic (ROC) _{curves. The corrections could also be determined using linear or non-linear} calibration factors derived from repeated scans of the same patient (or phantom object) on different scanner set ups. Alternatively, a neural network based approach could be used. _{It is envisaged that already established biomarkers of vascular inflammation and} biological states such as fibrosis, oedema,vascularity, lipolysis and/or adipogenesis; and established risk factors can be combined with the output values of the invention to provide a more complete risk score for a given subject. Accordingly, the computer implemented method according to the invention may further comprise using other biomarkers in step (c) of claim 1. _{In a preferred embodiment, step (b) and/or (c) of the methods of the invention} comprises determining one or more of the following vessel features: (_{vii) calcium index (Calcium-i);} _{(viii) perivascular water index (PVWi);} _{(ix) epicardial adipose tissue volume (EpAT-vol);} _{(x) fat attenuation index of epicardial adipose tissue (FAIEpAT);} _{(xi) fat attenuation index of perivascular adipose tissue (FAIPVAT);} _{(xii) fibrous plaque index (FPi);} _{(xiii) perivascular water index (PVWi);} _{(xiv) volumetric perivascular characterisation index (VPCI);} _{(xv) the presence, volume or radiomic profile of plaque;} _{(xvi) the presence, volume or radiomic profile of high-risk plaque;} _{(xvii) the presence, volume or radiomic profile of low-attenuation plaque;} _{(xviii) the presence, volume or radiomic profile of lipid-rich plaque;} _{(xix) the presence, volume or radiomic profile of fibrous plaque;} _{(xx) the presence, volume or radiomic profile of non-calcified plaque;} _{(xxi) the presence, volume or radiomic profile of calcified plaque; and} _{(xxii) Vessel volume, diameter, cross-sectional area, surface area, length,} location or remodelling. The non-plaque, vessel specific features (e.g. (i), (ii), (v), (vi) and (viii)) may be used in step (b) to determine the presence of disease and/or vascular state. For instance, one or more of: inflammation, oedema, fibrosis, vascularity, lipolysis and/or adipogenesis. The one or more vessel features could be determined for one or more vessels. _{The computed risk may comprise determining the number of vessels or segments} _{thereof having the one or more vessel features present or above or below a} specified threshold. Step (b) or (c) may further comprise determining one or more of the following plaque features: (_{xxiii) Plaque density;} _{(xxiv) Composition;} _{(xxv) Calcification;} _{(xxvi) Radiodensity;} _{(xxvii) Location;} _{(xxviii) Volume;} _{(xxix) surface area;} _{(xxx) geometry;} _{(xxxi) heterogeneity;} _{(xxxii) diffusivity; and} _{(xxxiii) ratio between volume and surface area.} Step (c) of the method may also comprise taking into account one or more of the following risk factors or characteristics of the subject: (_{xxxiv) age of the subject;} _{(xxxv) sex/gender of the subject;} _{(xxxvi) race/ethnicity of the subject;} _{(xxxvii) coronary calcium;} _{(xxxviii) hypertension;} _{(xxxix) hyperlipidemia/hypercholesterolemia;} _{(xl) diabetes mellitus;} _{(xli) presence of coronary artery disease;} _{(xlii) smoking;} _{(xliii) family history of heart disease; and} _{(xliv) genetic status.} These definitions of these biomarkers are known to the skilled person, and are described extensively in WO 2018/078395 A1, the contents of which are herein incorporated by reference. _{The output value may be compared to an earlier scan of the same subject to} _{compute a local change in status of disease and/or vascular status e.g.} inflammation, oedema, fibrosis, vascularity, lipolysis and/or adipogenesis. The _{methods of the invention, when the number of carotid arteries is analysed, may} also be used to non-invasively monitor carotid plaques. The methods of the _{invention also enable the quantification of disease and/or vascular state e.g. one} or more of vascular inflammation, fibrosis, oedema, vascularity, lipolysis and adipogenesis; and guide pharmacological treatment decisions. In a preferred embodiment, the output value is used to predict the cardiac mortality risk, risk of a subject suffering a cardiovascular event, and risk of disease development, progression or regression. In embodiments, the output in step (c) may be generated using a trained machine learning or statistical model, wherein the input to the model comprises a metric _{indicative of the number of vessels or segments thereof having disease and/or} vascular state (e.g. inflammation, oedema, fibrosis, vascularity, lipolysis and/or _{adipogenesis). The model is typically a regression or classifier model such as a} Cox Proportional-Hazards model. However, other suitable machine learning or _{statistical models, such as artificial neural networks, may be used as would be} understood by the person skilled in the art. The metric indicative of the number vessels or segments thereof having disease _{and/or (adverse) vascular state may be a number of blood vessels (or segments} _{thereof) that are determined to have disease and/or vascular state in step (b) of} the method. In other embodiments, the metric indicative of the number of vessels _{or segments thereof having disease and/or (adverse) vascular state may be} _{disease and/or vascular state metric (e.g. an inflammation, oedema, fibrosis,} vascularity, lipolysis and/or adipogenesis metric) of each vessel or segment. Preferably, the metric is an inflammation metric and comprises a FAI score of the perivascular adipose tissue surrounding each blood vessel or segment. _{Preferably, the input to the model further comprises one or more additional risk} _{metrics, preferably wherein the one or more additional risk metrics comprise at} least one of: (_{i) one or more plaque features of the blood vessel or segments thereof;} _{(ii) one or more risk factors or characteristics of the subject.} The use of additional risk metrics in the model for generating an output value that indicates the subject’s presence or risk of disease and/or (adverse) vascular state _{(e.g. risk of cardiac mortality or risk of suffering a cardiovascular event) may} _{advantageously provide a greater accuracy prediction of the subject’s risk.} _{Examples of plaque features that may be used as additional risk metrics include} the total volume of plaque; the volume of calcified plaque; the volume of non- calcified plaque; and/or the radiomic profile of the plaque voxels, as described above. Further plaque features that may be used as additional risk metrics include _{those listed at (xxiii) to (xxxiii) above.} _{The one or more risk factors or characteristics of the subject may comprise clinical} _{risk factors such as a smoking metric, a diabetes mellitus metric, or an age of the} subject. Examples of risk factors or characteristics of the subject include those _{listed at (xxxiv) to (xliv) above.} In some embodiments, the trained machine learning or statistic model used to generate the output value in step (c) may be trained on one or more vessel features of the blood vessels. The vessel features may include the lumen diameter in each portion of a vessel; the lumen volume of each portion of the vessel; the percentage of fractional myocardial mass; or the total length of vessel distal to each portion of the vessel. In this way, the relative significance of each _{vessel or segment thereof may be taken into account when determining the risk} _{score. Further examples of such vessel features include those listed at (i) to (xxii)} above. _{The methods of the invention are computer implemented, and therefore require a} processing system. Preferably, the entirety of the methods are computer- implemented. Accordingly, the methods of the invention can proceed automatically without manual intervention when given input medical imaging data. For the interests of quality control, however, it is envisaged that a trained operator may check and adjust particular parameters to ensure compliance with established clinical practices. To this end, the methods of the invention may be implemented automatically using dedicated software providing a rapid, non-invasive estimation of an individual’s _{risk status and likelihood of adverse events, and guide clinical decision making.} _{In a second aspect, the invention provides a computer program product} comprising executable instructions which, when executed by one or more processors, cause the one or more processors to perform the method of the first aspect. _{In a third aspect, there is provided a system comprising one or more processors} configured to perform the method of the first aspect. Further disclosed is a non-transitory computer readable medium comprising executable instructions which, when executed by one or more processors, cause the one or more processors to perform the method according to the first aspect. _{Figure 3 schematically illustrates an example of a system suitable for} implementing embodiments of the method. The system 1130 comprises at least one server 1110 which is in communication with a reference data store 1120. The server may also be in communication with other hardware which may be operated by a healthcare professional, for example over a communications network 1140. In certain embodiments the server may obtain, for example using from the reference data store, pre-determined threshold values which may be corrected for _{biological and technical factors. The server may then provide a corrected output} value according to the methods described herein that determine the presence or _{risk of disease and/or (adverse) vascular state, to provide a useful clinical picture} of the subject’s current vessel health and risk of suffering adverse cardiac events. An example of a suitable server 1110 is shown in Figure 4. In this example, the server includes at least one microprocessor 1200, a memory 1201, an optional input/output device 1202, such as a keyboard and/or display, and an external interface 1203, interconnected via a bus 1204 as shown. In this example the external interface 1203 can be utilised for connecting the server 1110 to peripheral devices, such as the communications networks 1140, reference data store 1120, other storage devices, or the like. _{Although a single external interface 1203 is shown, this is for the purpose of} _{example only, and in practice multiple interfaces using various methods (e.g.} _{Ethernet, serial, USB, wireless or the like) may be provided.} In use, the microprocessor 1200 executes instructions in the form of applications software stored in the memory 1201 to allow the required processes to be performed, including communicating with the reference data store 1120 in order to receive and process input data, and to provide a preferably corrected output _{score according to the methods described above. The applications software may} include one or more software modules, and may be executed in a suitable execution environment, such as an operating system environment, or the like. Accordingly, it will be appreciated that the server 1200 may be formed from any suitable processing system, such as a suitably programmed client device, PC, web server, network server, or the like. In one particular example, the server 1200 is a standard processing system such as an Intel Architecture based processing _{system, which executes software applications stored on non- volatile (e.g., hard} disk) storage, although this is not essential. However, it will also be understood that the processing system could be any electronic processing device such as a microprocessor, microchip processor, logic gate configuration, firmware optionally associated with implementing logic such as an FPGA (Field Programmable Gate Array), or any other electronic device, system or arrangement. Accordingly, whilst the term server is used, this is for the purpose of example only and is not intended to be limiting. Whilst the server 1200 is a shown as a single entity, it will be appreciated that the server 1200 can be distributed over a number of geographically separate locations, for example by using processing systems and/or databases 1201 that are provided as part of a cloud based environment. Thus, the above described arrangement is not essential and other suitable configurations could be used. In a preferred embodiment, the processing systems are cloud based. In accordance with a fourth aspect, there is provided a computer-implemented method for determining the presence or risk of disease and/or vascular state in a subject, comprising: a_{ccessing a metric indicative of the number of blood vessels, or segments} _{thereof, having disease and/or (adverse) vascular state, from two or more blood} _{vessels of the subject; and} _{inputting the metric indicative of the number of blood vessels or segments} _{thereof having disease and/or (adverse) vascular state, into a trained machine} learning or statistical model to determining the presence or risk of disease and/or vascular state in a subject. Preferably, the metric is an inflammation metric, and the metric is indicative of the number of inflamed blood vessels. In a preferred embodiment, the metric is indicative of the number of inflamed blood vessels. The number of inflamed segments of these blood vessels may be used in combination with the number of inflamed blood vessels. _{The fourth aspect of the invention therefore advantageously provides an output} value determining the presence or risk of disease and/or vascular state in a subject. This value may be a classification (e.g. “high”, “medium” or “low”), or a continuous probability. This, in embodiments, output of the trained machine learning or statistical model is a classification or probability of disease and/or vascular state in a subject. The prediction may be used in combination with other metrics to determine a subject’s risk of a cardiac event. In some embodiments, the model may be used in step (c) of the methods described in the first aspect of the invention. _{The metric indicative of the number of blood vessels or segments thereof having} disease and/or vascular state (e.g. inflammation, oedema, fibrosis, vascularity, _{lipolysis and/or adipogenesis) may be a number of blood vessels (or segments)} that are determined as having disease and/or vascular state. The number of blood _{vessels or segments thereof having disease and/or vascular state, may be} determined using the techniques described herein. Alternatively or additionally, the metric indicative of the number of vessels or segments thereof having disease and/or (adverse) vascular state, may be an _{inflammation, oedema, fibrosis and/or vascularity metric of each of two or more} _{blood vessels, or segments thereof, of the subject. By using an inflammation} metric of two or more blood vessels (or segments) of the subject, the model provides an improved prediction of the cardiac mortality risk or risk of the subject suffering a cardiovascular event, compared to previously known methods in which _{the highest metric (e.g. FAI score) of any given vessel is used in a model to predict} cardiovascular risk. Preferably, the inflammation metric comprises a fat attenuation index (FAI) score of perivascular adipose tissue. This is typically derived from the radiodensity value _{of the perivascular tissue surrounding each blood vessel or segment thereof, for} example using techniques known in the art. Preferably, the method further comprises: accessing one or more additional risk metrics of the subject; and inputting the one or more additional risk metrics into the trained machine learning or statistical model together with the metric indicative _{of the number of vessels or segments thereof having disease and/or (adverse)} _{vascular state, to determine the presence or risk of disease and/or vascular state} in a subject. This may advantageously improve performance (e.g. higher accuracy predictions) of the model. _{Preferably, the one or more additional risk metrics comprises one or more plaque} features. Examples of plaque features that may be used as additional risk metrics include the total volume of plaque; the volume of calcified plaque; the volume of non-calcified plaque; and/or the radiomic profile of the plaque voxels, as described above. Further plaque features that may be used as additional risk metrics include _{those listed at (xxiii) to (xxxiii) above.} Preferably, the one or more additional risk metrics may comprise one or more risk _{factors or characteristics of the subject. These may be clinical risk factors such} _{as a smoking metric, a diabetes mellitus metric, or an age of the subject.} _{Examples of risk factors or characteristics of the subject include those listed at} _{(xxxiv) to (xliv) above.} Preferably, the input to the trained machine learning or statistical model may further comprise one or more vessel features of the blood vessels. The vessel features may include the lumen diameter in each portion of a vessel; the lumen volume of each portion of the vessel; the percentage of fractional myocardial mass; or the total length of vessel distal to each portion of the vessel. In this way, the relative significance of each vessel or segment thereof may be taken into account when determining the risk score. Further examples of such vessel _{features include those listed at (i) to (xxiii) above.} Preferably, the machine learning or statistical model is trained by: generating a reference dataset that comprises: (i) a plurality of reference metrics indicative of the number of b_{lood vessels, or segments thereof having disease and/or (adverse)} vascular state (preferably selected from: inflammation, oedema, fibrosis, _{vascularity, lipolysis and/or adipogenesis, most preferably inflammation;} _{from two or more blood vessels of the subject; and} _{(ii) a plurality of reference classifications of a cardiac mortality} risk or risk of a subject suffering a cardiovascular event associated with the reference metrics indicative of the number of blood vessels or s_{egments thereof having disease and/or (adverse) vascular state; and} training the machine learning or statistical model using the reference dataset to learn a relationship between the metrics indicative of the number of _{blood vessels or segments thereof having disease and/or (adverse) vascular} state, and the classifications of the presence or risk of disease and/or vascular state in a subject. The reference dataset comprises a plurality of reference classifications of the presence or risk of disease and/or vascular state in a subject. Preferably, the _{reference classifications may be in the form of adverse cardiac or cardiovascular} _{events recorded in subjects within a predetermined time period. Thus, the} predictions generated by the model may be indicative of the presence or risk of disease and/or (adverse) vascular state e.g. cardiac mortality risk or risk of a _{subject suffering a cardiovascular event, over the corresponding predetermined} time period. The time period may typically be a number of years, for example between 1-10 years. In a fifth aspect of the invention there is provided a method of training a machine learning or statistical model for determining the presence or risk of disease and/or vascular state in a subject, comprising: generating a reference dataset that comprises: (i) a plurality of reference metrics indicative of the number of b_{lood vessels, or segments thereof, having disease and/or (adverse)} vascular state, preferably selected from: inflammation, oedema, fibrosis, vascularity, lipolysis and/or adipogenesis; from two or more blood vessels o_{f the subject; and} _{(ii) a plurality of reference classifications of a cardiac mortality} risk or risk of a subject suffering a cardiovascular event associated with t_{he reference metrics indicative of the number of blood vessels or} _{segments thereof having disease and/or (adverse) vascular state; and} training the machine learning or statistical model using the reference _{dataset to learn a relationship between the metrics indicative of the number of} _{blood vessels or segments thereof having disease and/or (adverse) vascular} _{state, and the classifications of the presence or risk of disease and/or vascular} state in a subject. Preferably, the reference dataset further comprises one or more reference additional risk metrics of the subject. In this way, the model may advantageously be trained to learn a relationship between the metrics indicative of the number of _{blood vessels or segments thereof having disease and/or (adverse) vascular} _{state, the additional risk metric(s), and the classifications of the presence or risk} of disease and/or vascular state in a subject. Further disclosed is a machine learning or statistical model trained using the method according to the fifth aspect of the invention. Preferably, the machine learning or statistical model is a Cox Proportional- Hazards regression model. However, it is envisaged that any suitable classifier or regression model could be used. In other examples, the machine learning or statistical model may be or comprise an artificial neural network. Further disclosed is a system for predicting the cardiac mortality risk or risk of a patient suffering a cardiovascular event, comprising at least one processor in communication with at least one memory device, the at least one memory device having stored thereon instructions for causing the at least one processor to perform a method according to the fourth or fifth aspects. The preferred features as described for the first aspect of the invention are also preferred in the context of all other aspects of the invention. EXAMPLES Method Consecutive patients undergoing clinically indicated CCTA in 8 UK hospitals (n=40,091) were followed for Major Adverse Cardiac Events (MACE; myocardial infarction, new onset heart failure, or cardiac death) for a median(IQR) of 2·7(1·4- 5·3)years. The prognostic value of FAI Score in the presence and absence of obstructive CAD was evaluated in 3,393 consecutive patients from the 2 hospitals with the longest follow up [7·7(6·4-9·1)years]. An artificial intelligence-enhanced cardiac risk prediction algorithm (AI-Risk), that integrates FAI Score, coronary plaque metrics and clinical risk factors, was then evaluated in this population. The study encompassed 3 aims: (A) To evaluate the risk profile and event rates among patients undergoing CCTA as part of routine clinical care in the UK National Healthcare system (NHS); (B) to test the hypothesis that coronary arterial inflammation (measured using the perivascular FAI Score in any coronary artery) drives cardiac mortality or MACE in patients with/without CAD and (C) to externally validate the performance of the previously trained AI-Risk prognostic algorithm and the related AI-Risk Classification system in a UK population. The first objective was assessed in a large longitudinal cohort (n=40,091, Cohort A) while the other two objectives were assessed in a nested longitudinal study with a longer follow up (n=3393, Cohort B). Cohort A (ORFAN study): Understanding the risk profile of people undergoing CCTA This analysis was performed within the ORFAN study (NCT05169333 and https://oxhvf.com/the-orfan-study/), and included 40,091 consecutive patients undergoing CCTA as part of routine clinical care in 8 Hospitals in the UK (Oxford University Hospitals, Royal United Hospital Bath, Royal Papworth Hospital, Royal Brompton Hospital, Harefield Hospital, Leicester University Hospital, Milton Keynes Hospital, Leeds Teaching Hospitals) between 2011-2021. This ethnically diverse cohort represents the UK population (Table 1). Obstructive CAD on CCTA was defined as ≥50% stenosis of left main stem or ≥70% stenosis of any of the three major epicardial coronary arteries in accordance to SCCT/ACC/ACR/NASCI consensus. Coronary Calcium Score (CCS) was extracted from the clinical reports when a non-contrast scan was performed. Local databases were constructed based on electronic patient records within each hospital, and the clinical reporting was performed locally by trained clinicians. The CCTA scans were then transferred to the ORFAN study core lab, at the Acute Multidisciplinary Imaging and Interventional Centre (AMIIC) of the University of Oxford, using a General Data Protection Regulation (GDPR)-compliant gateway (CIMAR gateway, provided by Caristo Diagnostics Ltd). Patient demographics and clinical outcomes data were collected via local resources and nationwide databases (NHS Digital and the National Institute of Cardiovascular Outcomes Research [NICOR]) using ICD10 codes, and the study population was followed up prospectively for a median of 2·7 years (IQR 1·4-5·3). The study design, patient selection process and data linkage _{approach are presented in Figure 6. The ORFAN study was approved by the} Oxfordshire Research Ethics Committee (REC 15/SC/0545) and the UK Confidentiality Advisory Group (20/CAG/0157). Cohort B: Validating FAI Score, the AI-Risk prognostic algorithm and the AI- Risk Classification system To validate the long-term prognostic value of FAI Score and the performance of the AI-Risk algorithm, a nested cohort was designed within the ORFAN population, to include individuals who underwent CCTA in the two hospitals with the longest follow-up available (Royal Brompton and Harefield Hospitals). This cohort included 3,393 consecutive unselected patients undergoing clinically indicated _{CCTA, between the years 2010-2015 (Table 1 and Figure 6). Patients referred for} the evaluation of congenital heart disease or heart transplantation were excluded. These patients were followed for a median of 7·7 (interquartile range [IQR] 6·4- 9·1) years via data-linkage with nationwide databases (see below) for incident MACE (myocardial infarction, new heart failure, and cardiac mortality) and cardiac _{mortality as a separate endpoint. The CCTA scans were transferred to the ORFAN} core lab for analysis using the CaRi-Heart^® v2.0 device (Caristo Diagnostics Ltd, Oxford, UK) to generate the FAI Score for each coronary artery, the AI-Risk for the _{patient according to the quality standards regulating medical devices. The AI-Risk} Classification system categorises patients depending on their AI-risk and FAI Score as described below. The extent and severity of CAD was assessed by trained personnel in the ORFAN study core lab, by using the Coronary Artery _{Disease Reporting and Data System (CADRADS 2.0). Clinical reports of the CCTA} were obtained and cross-referenced with the core lab reports as an internal quality check of the study core lab’s plaque assessment. The results of the FAI Scores and AI-Risk, as well as the AI-Risk Classification were included into a database, which was locked before it was merged with the outcomes database for statistical analysis. QRISK3 was calculated using age, sex, ethnicity, smoking, diabetes, _{family history, chronic kidney disease, atrial fibrillation, blood pressure treatment,} migraines, rheumatoid arthritis, systemic lupus erythematosus, severe mental illness, antipsychotic medication, steroid tablets, body mass index and lipid profile [https://qrisk.org/]. The QRISK3 model was originally developed using the NHS Digital data from the UK population between 1998-2015. Procedures FAI Score and AI-Risk were computed using the CaRi-Heart^® V2.0 medical device (Caristo Diagnostics Ltd, Oxford, UK). Descriptions of the algorithms used in the device were presented previously,^1-3. Briefly, The CCTA scans are uploaded into a medical device called CaRi-Heart. A deep learning model performs the segmentation of the arterial wall and the perivascular space and calculates the FAI Score. FAI Score assesses the degree of inflammation in each of the three main epicardial coronary arteries (right coronary artery (RCA), left anterior descending coronary artery (LAD) and the left circumflex artery (LCx)), and is derived using a proprietary algorithm that incorporates FAI with adjustments for age, sex, scan technical parameters, biological and anatomical factors, as _{previously described.3 The readout presents age- and sex-specific nomograms for} clinical use, based on its regulatory label and the 2023 European Society of Cardiology (ESC) Clinical Consensus Statement on using PVAT imaging for risk stratification.⁴ A patient is considered as exposed to high inflammatory risk if their FAI Score in the LAD or the RCA is above the 75^th percentile for the patient’s age and sex, or above the 95^th percentile in the LCx while they are considered to be of very high-risk if the FAI Score is above the 90^th percentile in either the LAD or the RCA, as previously reported ^1,3 and adopted in the recent ESC Clinical Consensus Statement.⁴ Previous studies have shown that FAI Score captures cardiovascular inflammatory risk and changes in response to risk-modifying treatments.^5,6 According to the technical file of the CaRi-Heart^® V2.0 medical device that performs these analyses, FAI Score has an intraclass correlation coefficient of 0·990 (p<0·001) for the LAD, 0·992 (p<0·001) for the LCx, and 0·980 (p<0·001) for the RCA, suggesting very low inter-observer variability. The FAI Score of the most inflamed artery is then incorporated into a prognostic model together with traditional clinical risk factors (diabetes, smoking, hyperlipidaemia, and hypertension) and plaque burden (modified Duke CAD index, an angiographic score integrating proximal CAD, plaque extent, and left main disease)⁷ to generate the 8-year % risk of the individual patient for a fatal cardiac event (AI-Risk algorithm).³ This prognostic model was trained in the USA population of the CRISP-CT study and validated in a European cohort of nearly two thousand patients, being calibrated to predict the 8-years risk for cardiac mortality. The current study examines the generalisability and validity of the algorithm, in an independent cohort from a different geographical area and a different demographic profile.³ Time-limited risk prediction models underestimate risk in young individuals, and this limitation is addressed by the AI-Risk Classification, which takes into account both the FAI Score (reflecting the disease inflammatory activity in the coronary arteries at the time of the CCTA scan) as well as the patient’s 8-year % risk for cardiac death (AI-Risk), as discussed in the recent ESC Clinical Consensus Statement.⁴ The AI-Risk Classification distributes patients into three risk categories based on their AI-Risk and FAI-Score, as follows:⁴ -_{Low/medium-risk category: AI-Risk <5% and FAI Score <75th percentile in} in the LCx; -_{High-risk category: AI-Risk 5% to <10% or FAI Score in the LAD/RCA} between 75^th and 90^th percentile or FAI Score in the LCx >95^th percentile; -_{Very high-risk category: AI-Risk ≥10% or FAI Score at LAD/RCA >90th} percentile. Finally, to evaluate the impact of AI-classification on clinical decision-making, a prospective real-world evaluation survey was conducted in 4 NHS Hospitals, that involved 744 consecutive patients undergoing CCTA for investigation of chest pain as described below. _{Further information on the algorithms used in the CaRi-Heart® V2.0 device can} be found in the references 8-10 and 15. Statistical analysis Patient baseline characteristics were compared using Pearson’s χ² or Fisher’s exact test for categorical variables, and t-test and ANOVA (three groups) for continuous variables, as appropriate. Individual follow-up time was calculated from the date of the CCTA until the date of occurrence of first MACE or the last date of data extraction (March 31, 2021). Probabilities for any event for the first time since CCTA were plotted using the Kaplan-Meier failure curves. Multivariable cox-regression model was fitted to estimate the hazard rates, hazard ratios (HR) and the 95% confidence intervals (CI) for obstructive CAD, FAI Scores, AI-Risk _{(as continuous variable) and AI-Risk Classification (categorical variable) on} clinical outcomes including MACE and cardiac mortality. The HR for FAI Scores _{(already adjusted for age-, sex- and technical parameters) were adjusted for} clinical risk factors (hypertension, diabetes, smoking, hyperlipidemia) the extent of CAD using the CAD-RADS 2.0 classification system¹¹, medications and prior coronary revascularisation. Schoenfeld residuals plots visually assessed proportional hazards assumptions. There were none to minimal violations of the assumptions on FAI Score, AI-Risk or AI-Risk Classification for all the events. Patients who contributed only one time- point were included in the analysis and assigned a follow-up time of one day. Missing values were imputed for smoking status using MICE package in R with CART (classification and regression trees) method as stated in the Results section. Imputation for smoking status was performed based on the demographics (age, sex, ethnicity) as well as smoking related diseases recorded at the end of the follow up period in 2022. This includes cancer, respiratory, circulatory diseases, mental health conditions and other diseases as previously described.¹⁷ To get a single value out of the 20 imputed datasets, a 10-fold validation cross was used to find the best performing method out of k-nearest neighbour, naїve 5 Bayes, and CART. The CART method was preferred due to its high accuracy (0·82) compared to the rest. The output of the AI-Risk algorithm was compared to the baseline model of QRISK3 to understand its incremental prognostic value in this patient population. 10 Improvement in discrimination was assessed by comparing the time-dependent c- statistic of the two models¹⁸, as well as by calculating the net reclassification improvement (continuous NRI) and integrated discrimination improvement (IDI) (95% CI calculated using bootstrapping with 200 replications) between the two models.¹⁹ All analyses were done using a 10 year horizon. Calibration was 15 assessed by fitting Kaplan-Meier estimates with the mean predicted survival probabilities across different follow-up times. Statistical tests were performed using Stata 18.0 (StataCorp LP, College Station, TX) and the R environment (R 4.0.2, The R Foundation for Statistical Computing, 20 http://www.R-project.org) using R studio (version 4.0.2) and the following packages: rms, survival, riskRegression, survIDINRI, timeROC, and survivalROC. All tests were two-sided and values of P<0·05 were considered statistically significant. 25 Results T_{he clinical characteristics of the study participants are presented in Table 1.} _{Characteristics Cohort A} Cohort B (N=40,091) (N=3,393) Demographics _{Median age (IQR), years 59 (50-70) 62 (50-73)} _{Male, n (%) 21,366 (53·3) 1,914 (56·4)} Ethnicity, n (%) W_{hite 31,075 (77·5) 2,599 (76.6)} _{Asian 3,697 (9·2) 279 (8.2)} _{Black 1,002 (2·5) 100 (2.9)} _{Other groups 2,842 (7·1) 334 (9.8)} _{Unknown 1,475 (3·7) 81 (2.4)} _{Median follow up years (IQR) 2·7 (1·4-5·3) 7·7 (6·4-9·1)} Cardiovascular risk factors, n (%) _{Hypertension 16,963 (42·3) 2,060 (60·7)} _{Hyperlipidaemia 10,238 (25·5) 1,340 (39·5)} _{Diabetes mellitus 7,308 (18·2) 552 (16·3)} _{Smoking 4,802 (12·0) 608 (17·9)} QRISK3 score Low/medium risk _{26,167 (65·3) 2073 (61·3)} (<10%) H_{igh risk (10-19%) 10,281 (25·6) 1003 (29·7)} _{Very high risk (≥20%) 3,643 (9·1) 303 (9·0)} _{Coronary Calcium Score ≥300 † 2,012 (20·3) 397 (30·5)} _{History of myocardial infarction 1,981 (4.9) 203 (5·9)} _{History of PCI 1,656 (4.1) 201 (5·9)} _{History of CABG 733 (1.8) 28 (0·8)} Events/procedures after CCTA, n (%) _{MACE 4,307 (10·7) 706 (20·8)} _{Non-fatal myocardial infarction 1,898 (4·7) 297 (8·8)} _{New heart failure 1,727 (4·3) 313 (9·2)} _{Stroke 668 (1·7) 110 (3·2)} _{Cardiac death 1,754 (4·4) 339 (10·0)} _{Non-cardiac death 3,501 (8·7) 662 (19·5)} _{All-cause death 5,255 (13·1) 1,001 (29·5)} _{PCI 3,116 (7·8) 388 (11·4)} _{CABG 1,009 (2·5) 139 (4·1)} Medications, n(%) _{Antiplatelets 15,839 (43·1) 1,126 (42·3)} _{Warfarin 2,675 (7·3) 272 (10·2)} _{Beta blockers 17,329 (47·1) 1,191 (44·8)} _{Calcium channel blockers 11,770 (32·0) 824 (31·0)} _{Nitrates 9,153 (24·9) 411 (15·5)} _{Statins 22,844 (62·1) 1,716 (64·5)} _{ACE inhibitors 12,379 (33·6) 892 (33·5)} _{Angiotensin receptor blockers 6,993 (19·0) 579 (21·8)} _{Diuretics 12,939 (35·2) 1,024 (38·5)} _{Digoxin 1,261 (3·4) 118 (4·4)} _{Insulin 1,610 (4·4) 120 (4·5)} _{Oral hypoglycaemics 5,508 (15·0) 455 (17·1)} _{Direct oral anticoagulant 6,308 (17·1) 510 (19·2)} Table 1: Cardiovascular risk profile and outcomes in the presence or absence of obstructive CAD Within the whole population in Cohort A (ORFAN Study, n=40,091), 9·1% were conventionally classified as very high risk (QRISK3≥20%) and 25·6% as high risk (QRISK3 between 10-19%). Only 19% of patients undergoing CCTA had obstructive CAD sufficient to require further investigations or interventions. The clinical characteristics of the patients with obstructive CAD are summarised in Table 2. _{Characteristics Overall} No obstructive Obstructive P-value (N = 40,091) CAD CAD (N = 32,533) (N = 7,558) Demographics _{Median age, years 59 (50-70) 59 (49-69) (53-73) <0·001} 64 _{Male, n (%) 21,366 (53·3) 16,560 (50·9) (63·6) <0·001} 4,806 _{Ethnicity, n (%) <0·001} _{White 31,075 (77·5) 25,110 (77·2)} (78·9) 5,965 A_{sian 3,697 (9·2) 2,976 (9·1)} (9·5) 721 B_{lack 1,002 (2·5) 863 (2·7)} (1·8) 139 O_{ther groups 2,842 (7·1) 2,294 (7·1)} (7·3) 548 U_{nknown 1,475 (3·7) 1,290 (4·0)} (2·4) 185 Cardiovascular risk factors, n (%) _{Hypertension 16,963 (42·3) 12,757 (39·2) (55·6) <0·001} 4,206 _{Hyperlipidaemia 10,238 (25·5) 7,298 (22·4) (38·9) <0·001} 2,940 _{Diabetes mellitus 7,308 (18·2) 5,497 (16·9) (24·0) <0·001} 1,811 _{Smoking 4,802 (12·0) 3,938 (12·1) (11·4) 0·105} 864 _{QRISK3 score <0·001} Low/medium _{26,167 (65·3) 22,285 (68·5) 3,882 (51·4)} risk (<10%) High risk (10- _{10,281 (25·6) 7,678 (23·6) 2,603 (34·4)} 19%) Very high risk _{3,643 (9·1) 2,570 (7·9) 1,073 (14·2)} (≥20%) Events/procedures after CCTA, n (%) _{MACE 4,307 (10·7) 2,857 (8·8) (19·2) <0·001} 1,450 _{Non-fatal MI 1,898 (4·7) 1,120 (3·4) (10·3) <0·001} 778 _{Heart failure 1,727 (4·3) 1,219 (3·7) (6·7) <0·001} 508 _{Stroke 668 (1·7) 485 (1·5) (2·4) <0·001} 183 _{Cardiac death 1,754 (4·4) 1,118 (3·4) (8·4) <0·001} 636 _{Non-cardiac death 3,501 (8·7) 2,628 (8·1) (11·6) <0·001} 873 _{All-cause death 5,255 (13·1) 3,746 (11·5) (20·0) <0·001} 1,509 _{PCI 3,116 (7·8) 1,832 (5·6) (17·0) <0·001} 1,284 _{CABG 1,009 (2·5) 461 (1·4) (7·3) <0·001} 548 Medications, n (%) _{Antiplatelets 15,839 (43·1) 11,564 (13·5) (63·3) <0·001} 4,275 _{Warfarin (7·3) 2,126 (7·1) (8·1) 0·003} 2,675 549 _{Beta blockers 17,329 (47·1) 13,267 (44·2) (60·2) <0·001} 4,062 Calcium channel _{11,770 (32·0) 9,376 (31·2) (35·5) <0·001} blockers 2,394 _{Nitrates 9,153 (24·9) 6,527 (21·7) (38·9) <0·001} 2,626 _{Statins 22,844 (62·1) 17,693 (58·9) (76·3) <0·001} 5,151 _{ACE inhibitors (33·6) 9,491 (31·6) (42·8) <0·001} 12,379 2,888 Angiotensin _{6,993 (19·0) 5,456 (18·2) (22·8) <0·001} receptor blockers 1,537 _{Diuretics (35·2) 30,044 (33·6) (42·3) <0·001} 12,939 2,857 _{Digoxin (3·4) 1,007 (3·4) (3·8) 0·119} 1,261 254 _{Insulin 1,610 (4·4) 1,139 (3·8) (7·0) <0·001} 471 Oral _{5,508 (15·0) 4,193 (14·0) (19·5) <0·001} hypoglycaemics 1,315 Direct oral _{6,308 (17·1) 5,015 (16·7) (19·2) <0·001} anticoagulant 1,293 Table 2 After adjusting for age, sex, cardiovascular risk factors, medications and history of MI or previous revascularisation, patients with obstructive CAD had significantly _{higher risk for both MACE and cardiac mortality (Figure 7), as well as myocardial} infarction (MI), new heart failure, ischaemic stroke and all-cause mortality.The results were similar after excluding patients with congenital heart disease or _{history of heart transplant. Importantly, the total number of cardiovascular events} in those without obstructive CAD was two-fold higher (n=1118 cardiac deaths and 2857 MACE in a population of 32,533) than the events observed among those with obstructive disease (n=636 cardiac deaths and 1450 MACE in a population of 7,558), during the 2·7-year median follow-up period. This highlights the unmet need to improve risk stratification and management of the population without obstructive CAD. The role of coronary inflammation in cardiovascular risk prediction During the median 7·7-year follow up in the cohort B, FAI Score predicted cardiac mortality and MACE in both, patients with or without obstructive CAD, an effect that was consistent across all coronary territories, i.e. the LAD ), LCx and RCA _{(Table 2), reflecting the total inflammatory cardiovascular risk. Given that clinical} cardiovascular risk factors contribute to atherogenesis at least partly by increasing vascular inflammation, the remaining residual inflammatory cardiovascular risk (beyond the clinical risk factors) was assessed by adjusting the respective HRs for cardiovascular risk factors (including hypertension, diabetes, hyperlipidaemia and smoking status) and the extent of coronary _{atherosclerosis present (CAD-RADS 2.0 classification)11 (Table 2). The results} _{showed that FAI Score captures the residual inflammatory risk of patients with or} without obstructive CAD, even after adjusting for risk factors and the extent of any non-obstructive atheroma. In a subgroup analysis among patients without prior MI or revascularisation (PCI or CABG), FAI Score in any artery remained predictive of cardiac mortality and MACE. In a further sensitivity analysis that included 1,300 patients with non-contrast CT scans also available in addition to the CCTA, FAI Score in any coronary artery remained highly significant in predicting cardiac _{mortality or MACE, even after adjusting for coronary calcium score (CCS), and} FAI Score was also predictive among patients with chronic inflammatory diseases. Furthermore, among patients with no or minimal atheroma on CCTA (CADRADS 2.0 score 0 / 1, n= 1,678), FAI Score in any coronary artery remained predictive of both cardiac mortality and MACE. Although the presence of one inflamed artery was enough to provide significant prognostic value independently from the vessel used for the measurement, an increase in the number of vessels with FAI Score >75^th percentile was related with a parallel increase in the risk for both cardiac mortality and MACE vs patients with all 3 arteries below the 25^th percentile, in both the presence and absence of obstructive CAD (Figure 1). Performance of AI-Risk prognostic model and the AI-Risk Classification The performance of the AI-Risk algorithm was validated in Cohort B. Both the calibration curve for AI-Risk as a continuous output variable (8-year % risk of cardiac mortality), as well as the AI-Risk Classification system (as three risk categories) showed excellent alignment between predicted and observed events in the overall population as well as in those without obstructive CAD (Figure 8). The AI-Risk algorithm appeared to underestimate risk in those with obstructive CAD, as the CCTA report triggered invasive coronary angiography and interventions (revascularisation and/or aggressive medical therapy) which are expected to modify the link between coronary inflammation at the time of the scan and cardiovascular events happening during the initial years after the test. Given that pre-specified clinical endpoints included only cardiac death, and MACE (MI, _{new HF and cardiac death), patients undergoing elective revascularisation} procedures after the CCTA scan were not censored, unless one of the study endpoints was met. This was more evident in the mid/low and high-risk classes, where adjustment of the model for calcified and non-calcified plaque volume had minimal impact on its performance (Figure 9). In the Cox regression model, the AI-Risk algorithm remained a significant and independent predictor of both cardiac mortality and MACE over a 10-year period (Table 2). Using the AI-Risk _{Classification system, patients in the very high-risk category had significantly} higher risk for both cardiac mortality and MACE compared to those in the low/medium-risk category, a finding replicated both in patients with and without obstructive CAD (Figure 10). Among patients with no or minimal atheroma (CADRADS 2.0=0/1), the very high AI-Risk class was associated with 9.5 times higher risk for cardiac mortality and 5.5 times higher risk of MACE, compared to the low/mid AI-Risk category (Figure 11). Compared to QRISK3, the AI-Risk Classification system significantly reclassified patients for both cardiac mortality [Net reclassification index (NRI) 0·38 (0·23- 0·45) P<0·0001 and integrated discrimination improvement (IDI) 0·028(0·014- 0·047) P<0·0001] and MACE [NRI 0·27 (0·091-0·32) P<0·0001 and IDI _{0·024(0·006-0·058) P<0.0001] over a 10 year horizon (Figure 12). Importantly, the} results were similar in the population without obstructive CAD, who are typically returned to primary care for further management. In the ROC analyses, the AUC for predicting cardiac mortality over a 10-year horizon using QRISK3 was 0·831 in the whole population, 0·786 in those with no obstructive CAD and 0·747 in those with obstructive CAD. The addition of CAD stenoses severity (CADRADS 2.0) to QRISK3 did not significantly improve prediction of cardiac mortality, with AUC 0·838 (p=0·36 against QRISK3) in the whole population, 0·788 (p=0·83) in those with no obstructive CAD, and 0·732 (p=0·51) in those with obstructive CAD. Adding AI-Risk (as continuous variable) to a baseline model that included CADRADS 2.0 and QRISK3 increased the AUC to 0·854 (p=7·7×10^-7 against QRISK3+CADRADS 2.0) in the whole population, 0·816 (p=0·0017) in those without obstructive CAD and 0·773 (p=8·9×10^-4) in those with obstructive CAD. Similarly, for prediction of MACE, QRISK3 had a good performance in this UK population (AUC 0·784 in the whole population, 0·731 in those with no obstructive CAD and 0·750 in those with obstructive CAD). These estimates did not improve significantly after addition of CADRADS 2.0, (AUC 0·789, p=0·38 in the whole population, 0·734 (p=0·73) in those without obstructive CAD and 0·731 (p=0·18) in those with obstructive CAD). The model that included QRISK3 and CADRADS 2.0 improved significantly by adding AI-Risk, to 0.805 (p=3·4x10^-8 against QRISK3+CADRADS 2.0) in the whole population, 0·748 _{(p=1·2x10-4) in those without obstructive CAD and 0·764 (p=6·6x10-2) in those with} obstructive CAD. The prognostic performance of FAI Score of each coronary vessel is presented in Table 3. FAI Score FAI Score FAI Score AI-Risk LAD (AUC) LCx (AUC) RCA (AUC) (AUC) Cardiac Mortality A_{ll 0.838 0.854 0.806 0·844} _{No obstructive CAD 0.809 0.818 0.769 0·798} _{Obstructive CAD 0.793 0.801 0.763 0·778} MACE A_{ll 0.779 0.795 0.756 0·797} _{No obstructive CAD 0.740 0.756 0.727 0·745} _{Obstructive CAD 0.745 0.741 0.703 0·761} _{Table 3 - Prognostic performance of FAI Score in each artery and AI-Risk in} patients with and without obstructive CAD LAD: Left anterior descending artery; LCX: Left circumflex artery; RCA: Right coronary artery; CAD: Coronary artery disease. In the prospective real-world evaluation survey, making the AI-Risk Classification available to the clinical care teams, resulted in a management change in 45% of patients (initiation of statin treatment (24%), increase in statin dosage (13%) and adding additional treatments beyond statins (8%), which included aspirin (2·4%), colchicine (8·3%) or icosapent-ethyl (0·4%)). Discussion This study demonstrates in a large outcomes cohort of individuals undergoing clinically indicated CCTA, that only a third of the future MACE happen among those patients with obstructive CAD, underlining the unmet need to develop tools that will identify the high-risk individuals in the absence of obstructive CAD. Measuring inflammation in any coronary artery by using the perivascular FAI Score revealed for the first time that a quarter of those individuals without obstructive disease had significantly elevated residual inflammatory risk that translated into a ten times higher risk for cardiac mortality or MACE over a 10-year period. The number of inflamed coronary vessels, identified by elevated FAI Score, exhibited an additive increase in the risk of cardiac mortality or MACE. An artificial intelligence-enhanced prognostic model (AI-Risk algorithm), that incorporates FAI Score, the extent of coronary atheroma (if any) as well as the patient’s traditional risk factors, was able to powerfully predict cardiovascular mortality and MACE over 10 years, both in the presence and absence of coronary atherosclerosis. Since the introduction of CCTA as a first-line investigation in the management of stable chest pain,^1-3,20 the global use of CCTA has increased sharply,⁴ with the majority of patients being referred back to primary care after exclusion of obstructive CAD.^4,5 This practice highlights an opportunity for healthcare systems to evaluate these individuals more deeply, to forestall future cardiovascular events in those without occlusive lesions. This study demonstrates that, among ~40,000 consecutive CCTAs performed in the UK as part of routine clinical practice, only 19% revealed obstructive CAD that guided further investigations/interventions. Although the presence of obstructive CAD was associated with a higher relative risk of adverse cardiovascular outcomes, in absolute numbers there were nearly twice as many cardiovascular events during the follow-up period in the much larger population without obstructive CAD compared to those with obstructive CAD. This observation supports the notion that acute coronary syndromes frequently result from the disruption of non-obstructive (presumably inflamed) atherosclerotic plaques.⁶ The use of current clinical risk prediction tools (e.g. QRISK3) in these patients is limited, as such models were developed in apparently healthy individuals and do not capture information such as CAD plaque burden _{and residual inflammatory risk. A robust risk prediction tool could therefore identify} the vulnerable patient with the inflamed coronary arteries, particularly in those without obstructive CAD. This approach would transform CCTA from a test to triage a minority of patients for further intervention to a prevention tool that guides management for all patients undergoing CCTA. Evidence from clinical trials suggests that anti-inflammatory treatments like statins²¹, colchicine²² or anti-IL1β²³ reduce cardiovascular events. Indeed, colchicine has been included in the ESC 2021 cardiovascular prevention guidelines,²⁴ and has received Food and Drug Administration (FDA) clearance with a broad cardiovascular risk reduction label.⁷ Given the potential unwanted actions of anti-inflammatory treatments, targeting treatments specifically to patients with coronary artery inflammation could improve their allocation more precisely than systemic markers such as hsCRP. Translational studies have discovered that inflammatory signals originating from the vascular wall activate perivascular lipolysis, triggering spatial changes in the perivascular adipose tissue composition.⁸ The FAI Score captures such findings on routine CCTA,^8-10 and it also appears to track the vascular effectiveness of anti-inflammatory treatments.^25,26 Current clinical guidelines recommend primary prevention for patients with a 10- year risk of ≥10% for MACE or ≥5% for fatal cardiac events.^1,2 However, the current 10-year prediction models (e.g. QRISK3) underestimate risk in young individuals and cannot capture the presence of non-obstructive CAD or the degree of coronary arterial inflammation. FAI Score identifies a large group of patients with elevated coronary artery inflammation, experiencing high relative risk for cardiac events, despite their low absolute 10-year risk (calculated by QRISK3) due to their young age. Integrating “disease activity” (FAI Score) with the CAD plaque burden and the patient’s risk factors provides a powerful risk assessment tool (AI-Risk algorithm).^9,10 The AI-Risk model validated in this study utilises the FAI Score of the artery with the highest inflammation, and retraining was not performed due to regulatory restrictions on the ‘locked’ model. However, the _{interesting findings of an additive impact of the number of inflamed coronary} vessels on risk prediction, together with emerging evidence on the prognostic value of plaque composition²⁷ and high risk plaque characterises²⁸, may justify retraining of AI-Risk model in the future, to include these additional metrics. The AI-Risk Classification system (that takes into account FAI Score and AI-Risk) provides a decision-making tool that enables meaningful risk stratification, informing risk-driven changes in management within the existing prevention _{guidelines. In this study, the AI-Risk Classification system identified the very-high} risk patients with significant risk for MACE and cardiac mortality, even among those with no or minimal coronary atheroma. By detecting coronary inflammation, the FAI Score identifies the disease activity, which precedes plaque formation and rupture, and could be involved in MI without obstructed coronary arteries.²⁹ This enables risk stratification in patients who would otherwise be reassured by the absence of obstructive CAD, but warrant consideration for individualised preventative management to modify residual inflammatory risk. This could be particularly useful in patients with autoimmune or chronic inflammatory diseases. On the other hand, understanding individualised inflammatory risk from CCTA could guide the intensification of statin or adjunctive anti-inflammatory treatments, beyond the indications listed in current clinical guidelines (which go beyond treating high cholesterol).¹ The study has some limitations. The performance of QRISK3 was higher than expected in predicting cardiac mortality or MACE, likely because QRISK3 was originally trained using data from the same source (NHS Digital), same population (UK population) and same time-period (pre-2017) as the ORFAN population used in this study. In contrast, AI-Risk was trained in US populations. This may explain the lack of incremental value of CADRADS 2.0 when added in a baseline model that included QRISK3, in predicting either MACE of cardiac mortality. However, this did not prevent a significant improvement of the model when adding AI-Risk, confirming that the current study represents true external validation of the AI-Risk in a cohort of different demographics from a different continent. In the population with obstructive CAD, although both the FAI Score and the AI-Risk classification accurately predict the true events in the very high-risk population from the first year after the scan, the survival curves between mid/low and high risk, only split after year 3. Indeed, patients diagnosed with obstructive CAD at the time of the CCTA undergo invasive angiograms and revascularisations or at least intensification of their medical therapy after the CCTA, which affects risk prediction based on CCTA analysis. Finally, plasma levels of inflammatory biomarkers such as high sensitivity C reactive protein (hs-CRP) were not available in the current cohort, so the incremental value of FAI Score in predicting cardiovascular outcomes beyond these biomarkers needs to be documented in future prospective outcome studies. Conclusions This study demonstrates that measuring coronary inflammation from routine CCTA captures cardiovascular inflammatory risk, particularly in patients without obstructive CAD and even in those without any visible plaque/coronary calcification. An AI-assisted risk prediction tool incorporating FAI Score, atherosclerotic plaque burden and patient risk factor profile provides clinically meaningful risk reclassification in patients undergoing routine CCTA that could guide the more precise use of preventative treatments including anti-inflammatory therapies. Improved prognostic model As discussed above, and as illustrated in Figure 1, an increase in the number of coronary vessels with FAI Score >75th percentile was related with a parallel increase in the risk for both cardiac mortality and MACE vs patients with all 3 coronary arteries below the 25th percentile, in both the presence and absence of obstructive CAD. Based on this identified relationship, we present a new prognostic model to predict the cardiac mortality risk or risk of a subject suffering _{a cardiovascular event, that is based on the number of inflamed vessels (or} _{segments thereof) identified using the methods described above.} The model may be trained on a reference dataset (e.g. “training dataset”) comprising the number of inflamed vessels (or segments thereof) of a patient, and _{the corresponding subject’s cardiac mortality risk or risk of the subject suffering a} _{cardiovascular event within a predetermined time period (e.g. an 8-year % risk as} discussed above) In some examples the predetermined time period may be a 10 year period whereby the model is trained to predict 10 year outcomes. In this way, _{the trained prognostic model may provide risk predictions based on the number of} _{inflamed vessels (or segments) input into the model. Alternatively or additionally} to vascular inflammation, the model may be trained on other vascular pathologies, such as oedema, fibrosis or vascularity. The model may be trained to provide a categorised risk prediction (e.g. “low/medium”, “high”, “very high”, or a continuous risk prediction (e.g. as a percentage risk of suffering a cardiac event within a predetermined time period). The machine learning or statistical model is a Cox regression model. However, it is envisaged that other regression or classifier models may be used. Training of the model may be performed using techniques known to the person skilled in the art. _{In some embodiments, the model may be further trained on one or more additional} _{risk metrics of the subject. In this way, the model may provide an improved} _{prediction (e.g. greater accuracy prediction) of the subject’s risk. The one or more} _{additional risk metrics may include one or more clinical risk factors and/or one or} more plaque or vessel features. In a preferred embodiment, the model is trained _{using a dataset of the following features (which may be categorical or continuous} metrics): _{- Number of inflamed vessels or segments thereof} _{- Modified Duke CAD index (represents atherosclerotic burden)} _{- A diabetes metric of the subject} _{- A smoking metric of the subject} _{- A hyperlipidaemia metric of the subject} _{- A hypertension metric of the subject.} In some embodiments, the model may be further trained on a FAI-score (typically weighted for age and gender). This may be the highest FAI score of the subject as discussed above, or may be a FAI score of each vessel (or segment thereof). References _{1. The National Institute for Health and Care Excellence (NICE). CVD risk} assessment and management (CG181). National Institute for Health and Care Excellence; 2020. _{2. Corrigendum to: 2021 ESC Guidelines on cardiovascular disease} prevention in clinical practice: Developed by the Task Force for cardiovascular disease prevention in clinical practice with representatives of the European Society of Cardiology and 12 medical societies With the special contribution of the _{European Association of Preventive Cardiology (EAPC). Eur Heart J 2022; 43(42):} 4468. _{3. The National Institute for Health and Care Excellence (NICE). NICE} Guidance for Stable Chest Pain Patients (CG95 & MTG32) to Appropriately Diagnose Patients with Suspected Coronary Artery Disease.2019. _{4. Dreisbach JG, Nicol ED, Roobottom CA, Padley S, Roditi G. Challenges} in delivering computed tomography coronary angiography as the first-line test for _{stable chest pain. Heart 2018; 104(11): 921-7.} _{5. SCOT-HEART investigators. CT coronary angiography in patients with} suspected angina due to coronary heart disease (SCOT-HEART): an open-label, _{parallel-group, multicentre trial. Lancet 2015; 385(9985): 2383-91.} _{6. Puchner SB, Liu T, Mayrhofer T, et al. High-risk plaque detected on} coronary CT angiography predicts acute coronary syndromes independent of significant stenosis in acute chest pain: results from the ROMICAT-II trial. J Am _{Coll Cardiol 2014; 64(7): 684-92.} _{7. Food and Drug Administration (FDA). LODOCO (Colchicine) Highlights of} prescribing information. In: Administration FaD, editor. US; 2023. _{8. Antonopoulos AS, Sanna F, Sabharwal N, et al. Detecting human} _{coronary inflammation by imaging perivascular fat. Sci Transl Med 2017; 9(398).} _{9. Oikonomou EK, Marwan M, Desai MY, et al. Non-invasive detection of} coronary inflammation using computed tomography and prediction of residual cardiovascular risk (the CRISP CT study): a post-hoc analysis of prospective _{outcome data. Lancet 2018; 392(10151): 929-39.} _{10. Oikonomou EK, Antonopoulos AS, Schottlander D, et al. Standardized} measurement of coronary inflammation using cardiovascular computed tomography: integration in clinical care as a prognostic medical device. _{Cardiovasc Res 2021; 117(13): 2677-90.} _{11. Cury RC, Leipsic J, Abbara S, et al. CAD-RADS 2.0 - 2022 Coronary} Artery Disease-Reporting and Data System: An Expert Consensus Document of the Society of Cardiovascular Computed Tomography (SCCT), the American College of Cardiology (ACC), the American College of Radiology (ACR), and the North America Society of Cardiovascular Imaging (NASCI). J Cardiovasc Comput _{Tomogr 2022; 16(6): 536-57.} _{12. Investigators S-H, Newby DE, Adamson PD, et al. Coronary CT} _{Angiography and 5-Year Risk of Myocardial Infarction. N Engl J Med 2018;} 379(10): 924-33. _{13. Medicines and Healthcare products Regulatory Agency (MHRA).} Regulating medical devices in the UK. In: Agency MaHpR, editor.; 2023. _{14. Hippisley-Cox J, Coupland C, Brindle P. Development and validation of} QRISK3 risk prediction algorithms to estimate future risk of cardiovascular _{disease: prospective cohort study. BMJ 2017; 357: j2099.} _{15. Antoniades C, Tousoulis D, Vavlukis M, et al. Perivascular adipose tissue} _{as a source of therapeutic targets and clinical biomarkers. Eur Heart J 2023.} _{16. Min JK, Shaw LJ, Devereux RB, et al. Prognostic value of multidetector} coronary computed tomographic angiography for prediction of all-cause mortality. _{J Am Coll Cardiol 2007; 50(12): 1161-70.} _{17. Public Health England. Consultation on proposed changes to the} calculation of smoking-attributable mortality and hospital admissions.2020. _{18. Blanche P, Dartigues JF, Jacqmin-Gadda H. Estimating and comparing} time-dependent areas under receiver operating characteristic curves for censored _{event times with competing risks. Stat Med 2013; 32(30): 5381-97.} _{19. Pencina MJ, D'Agostino RB, Sr., Steyerberg EW. Extensions of net} reclassification improvement calculations to measure usefulness of new _{biomarkers. Stat Med 2011; 30(1): 11-21.} _{20. Saraste A, Knuuti J. ESC 2019 guidelines for the diagnosis and} management of chronic coronary syndromes : Recommendations for _{cardiovascular imaging. Herz 2020; 45(5): 409-20.} _{21. Ridker PM, Danielson E, Fonseca FA, et al. Rosuvastatin to prevent} vascular events in men and women with elevated C-reactive protein. N Engl J Med 2008; 359(21): 2195-207. _{22. Nidorf SM, Fiolet ATL, Mosterd A, et al. Colchicine in Patients with Chronic} _{Coronary Disease. N Engl J Med 2020; 383(19): 1838-47.} _{23. Ridker PM, Everett BM, Thuren T, et al. Antiinflammatory Therapy with} _{Canakinumab for Atherosclerotic Disease. N Engl J Med 2017; 377(12): 1119-31.} _{24. Visseren FLJ, Mach F, Smulders YM, et al. 2021 ESC Guidelines on} _{cardiovascular disease prevention in clinical practice. Eur Heart J 2021; 42(34):} 3227-337. _{25. Elnabawi YA, Oikonomou EK, Dey AK, et al. Association of Biologic} Therapy With Coronary Inflammation in Patients With Psoriasis as Assessed by _{Perivascular Fat Attenuation Index. JAMA Cardiol 2019; 4(9): 885-91.} _{26. Oikonomou EK, Williams MC, Kotanidis CP, et al. A novel machine} learning-derived radiotranscriptomic signature of perivascular fat improves _{cardiac risk prediction using coronary CT angiography. Eur Heart J 2019; 40(43):} 3529-43. _{27. Tzolos E, Williams MC, McElhinney P, et al. Pericoronary Adipose Tissue} Attenuation, Low-Attenuation Plaque Burden, and 5-Year Risk of Myocardial _{Infarction. JACC Cardiovasc Imaging 2022; 15(6): 1078-88.} _{28. Goeller M, Achenbach S, Cadet S, et al. Pericoronary Adipose Tissue} Computed Tomography Attenuation and High-Risk Plaque Characteristics in Acute Coronary Syndrome Compared With Stable Coronary Artery Disease. _{JAMA Cardiol 2018; 3(9): 858-63.} _{29. Gaibazzi N, Martini C, Botti A, Pinazzi A, Bottazzi B, Palumbo AA.} Coronary Inflammation by Computed Tomography Pericoronary Fat Attenuation in _{MINOCA and Tako-Tsubo Syndrome. J Am Heart Assoc 2019; 8(17): e013235.} CLAIMS 1_{. A computer implemented method for determining the presence or risk of} disease and/or vascular state in a subject, said method comprising: (_{a) using data gathered from a computer tomography (CT) scan along a length} of two or more blood vessels to determine: (_{i) a value for the radiodensity and/or one or more radiomic features of} _{the perivascular tissue surrounding each blood vessel or segment} thereof; (_{b) using the value for the radiodensity and/or one or more radiomic features} of the perivascular tissue surrounding each blood vessel or segment t_{hereof to determine the presence of disease and/or vascular state in each} blood vessel, or segment thereof; and (_{c) generating an output value based on the number of vessels or segments} _{thereof with disease and/or vascular state that indicates the subject’s} presence or risk of disease and/or vascular state. _{2. The computer implemented method according to claim 1, wherein the} disease and/or vascular state is selected from one or more of: inflammation, oedema, fibrosis, vascularity, lipolysis and adipogenesis; preferably inflammation. _{3. The computer implemented method according to claim 1 or claim 2,} wherein step (a)(i) comprises determining a value for the radiodensity of the _{perivascular tissue surrounding each blood vessel or segment thereof, and using} this value in step (b). _{4. The computer implemented method according to any preceding claim,} wherein step (a)(i) comprises determining one or more radiomic features, preferably two or more radiomics features, of the perivascular tissue surrounding _{each blood vessels or segment thereof, and using this in step (b). The computer} implemented method according to any preceding claim, wherein the value for the radiodensity and/or one or more radiomic features in (a)(i) comprises determining the radiodensity and/or one or more radiomic features of perivascular adipose _{tissue surrounding each blood vessel or segment thereof.} _{5. The computer implemented method of any of the preceding claims,} wherein step (b) comprises comparing the value of disease and/or vascular state _{metric, preferably selected from an inflammation, oedema, fibrosis, vascularity,} _{lipolysis and/or lipogenesis metric, derived from the radiodensity value and/or one} _{or more radiomic features to a pre-determined cut-off value, or using the absolute} _{value, to determine whether each blood vessel or segment thereof has} inflammation, oedema, fibrosis, vascularity, lipolysis and/or adipogenesis. _{6. The computer implemented method according to claim 5, wherein the} _{metric is an inflammation metric, and comprises the fat attenuation index (FAI) of} perivascular adipose tissue (FAIPVAT). _{7. The computer implemented method according to any preceding claim,} wherein the blood vessels are coronary arteries. _{8. The computer implemented method according to claim 7, wherein the} coronary arteries are selected from two or more of, preferably each of: the right _{coronary artery, left anterior descending artery and left circumflex artery.} _{9. The computer implemented method according to any one of claims 1-6,} wherein the blood vessels are carotid arteries. _{10. The computer implemented method according to claim 9, wherein the} carotid arteries are the left common carotid artery and right common carotid artery. _{11. The computer implemented method according to claim 9, wherein the} carotid arteries are selected from two or more of: the left common carotid artery, the right common carotid artery, the left internal carotid artery, the right internal carotid artery, the left external carotid artery and the right external carotid artery. _{12. The computer implemented method according to any one of claims 9-11,} wherein the method is used to predict the risk of the subject suffering a stroke. _{13. The computer implemented method according to any one of claims 1-6,} _{wherein the blood vessel segments are segments or branches of the aorta.} _{14. The computer implemented method according to claim 13, wherein the} blood vessels segments are selected from the thoracic and/or abdominal segments of the aorta. _{15. The computer implemented method according to any preceding claim,} _{comprising determining the number of blood vessels or segments thereof that} have: disease and/or vascular state, preferably selected from: inflammation, _{oedema, fibrosis, vascularity, lipolysis and/or adipogenesis; and using this value} in step (c) of claim 1. _{16. The computer implemented method according to any preceding claim,} _{further comprising determining the length of each blood vessel or segments} _{thereof that has disease and/or vascular state, preferably selected from:} inflammation, oedema, fibrosis and/or vascularity; and using this value in step (c) of claim 1. _{17. The computer implemented method according to any preceding claim,} further comprising determining the total length of blood vessel or segments thereof that has disease and/or vascular state, preferably selected from: inflammation, _{oedema, fibrosis, vascularity, lipolysis and/or adipogenesis; across each of the} _{coronary arteries and/or each of the carotid arteries, and using this in step (c) of} claim 1. _{18. The computer implemented method according to any preceding claim,} further comprising quantifying the severity of disease and/or vascular state, preferably selected from: inflammation, oedema, fibrosis, vascularity, lipolysis _{and/or adipogenesis; in each length of inflamed blood vessel or segment thereof,} and using this value in step (c) of claim 1. _{19. The computer implemented method according to any preceding claim,} _{further comprising determining the total length of blood vessel or segment thereof} that has disease and/or vascular state; preferably selected from: inflammation,

Claims

oedema, fibrosis, vascularity, lipolysis and/or adipogenesis; across each of the coronary arteries and/or each of the carotid arteries weighted by: (_{i) the lumen diameter in each portion of a vessel;} _{(ii) the lumen volume of each portion of the vessel;} _{(iii) the percentage of fractional myocardial mass; or} _{(iv) the total length of vessel distal to each portion of the vessel; and} using this value in step (c) of claim 1. _{20. The computer implemented method according to any preceding claim,} _{further comprising quantifying the plaque burden in each blood vessel or in each} _{segment thereof, and using this value in step (c) of claim 1.} _{21. The computer implemented method according to claim 20, further} _{comprising determining the total length of blood vessel or segment thereof with} disease and/or vascular state, preferably inflammation, across each of the _{coronary arteries and/or each of the carotid arteries, weighted by the plaque} _{burden in each portion of the vessel or segment thereof, and using this value in} step (c) of claim 1. _{22. The computer implemented method according to claim 20 or claim 21,} wherein the plaque burden comprises determining: (_{i) the total volume of plaque;} _{(ii) the volume of calcified plaque; and/or} _{(iii) the volume of non-calcified plaque.} _{23. The computer implemented method according to any preceding claim,} wherein the threshold value is a distribution of values, preferably wherein the threshold value is vessel specific and/or specific to how proximal or distal along the vessel the value for the radiodensity of the perivascular tissue surrounding the blood vessel was derived. _{24. The computer implemented method according to any preceding claim,} wherein the pre-determined threshold value or output value is adjusted for biological factors, preferably wherein the biological factors are selected from one or more of the age of the subject, the gender of the subject, the ethnicity of the _{subject, the background adipocyte size or a surrogate for it, partial volume effects,} and the type of blood vessel. _{25. The computer implemented method according to any preceding claim,} wherein the pre-determined threshold value or output value is adjusted for technical factors, preferably wherein the technical factors are selected from one _{more of the tube voltage of the CT scanner, reconstruction algorithms, scanner} _{resolution, iodinated contrast agent, contrast type, injection rate, aortic contrast} opacification, left ventricular blood pool opacification, signal-to-noise, contrast-to- noise, milliamps, method of cardiac gating, single and multiple energy image acquisition, CT scanner type, heart rate, heart rhythm, or blood pressure. _{26. The computer implemented method according to any preceding claim,} wherein step (a)(i) comprises determining one or more radiomic features selected from: Short Run High Gray Level Emphasis, High Gray Level Emphasis, High Gray Level Run Emphasis, Autocorrelation, Sum Average, Joint Average, High Gray _{Level Zone Emphasis, Skewness, Skewness LLL, Kurtosis, 90th Percentile, 90th} _{Percentile LLL, Median LLL, Kurtosis LLL, Median, Run Entropy, Dependence} Entropy LLL, Dependence Entropy, Zone Entropy LLL, Run Entropy LLL, Mean _{LLL, Small Area Low Gray Level Emphasis, Low Gray Level Zone Emphasis,} Short Run Low Gray Level Emphasis, Low Gray Level Run Emphasis, Low Gray Level Emphasis, Small Dependence Low Gray Level Emphasis, Gray Level Variance LLL (GLSZM), Gray Level Variance (GLDM), Variance, Gray Level Variance (GLDM), Difference Variance LLL, Gray Level Variance LLL (GLRLM), Variance LLL, Gray Level Variance LLL (GLDM), Sum of Squares, Contrast LLL, Mean Absolute Deviation, Interquartile Range, Robust Mean Absolute Deviation, Long Run Low Gray Level Emphasis, Difference Variance, Gray Level Variance (GLSZM), Inverse Difference Moment Normalized, Mean Absolute Deviation LLL, _{Sum of Squares LLL, Contrast, Zone Entropy, Gray Level Non Uniformity} Normalized (GLRLM), Gray Level Non Uniformity Normalized LLL (GLRLM), Sum Entropy, Joint Energy, Entropy, Gray Level Non Uniformity Normalized (GLDM), Joint Energy, Gray Level Non Uniformity Normalized LLL (GLDM), Uniformity LLL, Sum Entropy LLL, Uniformity, Zone Entropy HHH, Size Zone Non Uniformity _{Normalized HHH, Small Area Emphasis HHH, Strength, Coarseness HLL,} Coarseness, Coarseness LHL, Coarseness LLL, Coarseness LLH, Coarseness HHH, Coarseness HLH, Coarseness HHL, Coarseness LHH, Cluster Tendency LLL, Cluster Tendency, Sum of Squares LLL, Mean Absolute Deviation LLL, Gray Level Variance LLL (GLDM), Variance LLL, Gray Level Variance LLL (GLRLM), Gray Level Variance (GLRLM), Robust Mean Absolute Deviation LLL, Gray Level Variance (GLDM), Variance, Mean Absolute Deviation, Cluster Prominence, Sum Entropy LLL, Interquartile Range LLL, Gray Level Variance LLL (GLSZM), Sum of Squares, Robust Mean Absolute Deviation, Sum Entropy, Interquartile Range, _{Cluster Prominence LLL, Entropy LLL, 10th Percentile LLL, 10th Percentile, Size} Zone Non Uniformity LLL, Dependence Non Uniformity HLL, Gray Level Non Uniformity HLL (GLSZM), Gray Level Non Uniformity (GLSZM), Run Length Non Uniformity HHL, Run Length Non Uniformity LHL, Dependence Non Uniformity LHL, Dependence Non Uniformity, Run Length Non Uniformity HLH, Busyness, Run Length Non Uniformity LLH, Dependence Non Uniformity LLH, Dependence Non Uniformity LLL, Size Zone Non Uniformity, Energy HLL, Run Length Non Uniformity LHH, Size Zone Non Uniformity HLL, Gray Level Non Uniformity LLH (GLSZM), Gray Level Non Uniformity LHL (GLSZM), Gray Level Non Uniformity LLL (GLSZM), Run Length Non Uniformity HLL, Gray Level Non Uniformity HLH (GLSZM), Gray Level Non Uniformity HHL (GLSZM), Run Length Non Uniformity, and Run Length Non Uniformity HHH; preferably wherein at least two radiomic features are selected. _{27. The computer implemented method according to any preceding claim,} further comprising using other biomarkers in step (c) of claim 1.

_{28. The computer implemented method according to any preceding claim,} wherein the output value is compared to an earlier scan of the same subject to _{compute a local change in status of disease or vascular state; preferably wherein} the disease or vascular state is selected from one or more of: inflammation, oedema, fibrosis, vascularity, lipolysis and adipogenesis. _{29. The computer implemented invention according to any preceding claim,} wherein the output value is used to determine disease progression or regression, or measure a change in disease or vascular status. _{30. The computer implemented invention according to any preceding claim,} wherein the method is used to predict risk of cardiac mortality, risk of a subject suffering a cardiovascular event, or risk of disease development. _{31. The computer implemented method according to any preceding claim,} wherein step (b) comprises determining the presence of inflammation, the number _{of inflamed blood vessels or segments thereof is the number of inflamed carotid} _{arteries or inflamed carotid artery segments, and the method is used to non-} invasively monitor carotid plaques. _{32. The computer implemented method according to any preceding claim,} wherein the method is used to quantify one or more of vascular inflammation, fibrosis, oedema, vascularity, lipolysis and adipogenesis. _{33. The computer implemented method of any of the preceding claims,} _{wherein the inflammation, oedema, fibrosis and/or vascularity metric is generated} using a trained machine learning or statistical model, wherein the input to the model comprises the value for the radiodensity and/or one or more radiomic _{features of each blood vessel or segments thereof.} _{34. The computer implemented method of claim 34, wherein the input to the} _{model further comprises at least one of:} (i) one or more technical factors; (ii) one or more biological factors of the subject.

_{36. The computer-implemented method of any of the preceding claims,} wherein the output value in step (c) is generated using a trained machine learning or statistical model, wherein the input to the model comprises a metric indicative _{of the number or vessels or segments thereof having inflammation, oedema,} fibrosis and/or vascularity. _{37. The computer-implemented method of claim 36, wherein the input to the} _{model further comprises one or more additional risk metrics, preferably wherein} _{the one or more additional risk metrics comprise at least one of:} _{(i) one or more plaque features of the blood vessel or segments thereof;} _{(ii) one or more risk factors or characteristics of the subject.} _{38. A computer program product comprising executable instructions which,} when executed by one or more processors, cause the one or more processors to perform the method of any preceding claim. _{39. A system comprising one or more processors configured to perform the} method of any preceding claim. _{40. A computer-implemented method for determining the presence or risk of} disease and/or vascular state in a subject, comprising: a_{ccessing a metric indicative of the number of blood vessels, or segments} _{thereof having disease and/or vascular state, from two or more blood vessels of} the subject; and i_{nputting the metric indicative of the number of blood vessels or segments} _{thereof having disease and/or vascular state, into a trained machine learning or} statistical model to determine the presence or risk of disease and/or vascular state. _{41. The method of claim 40, wherein the output of the trained machine} _{learning or statistical model is a classification or probability of the cardiac mortality} risk, risk of the subject suffering a cardiovascular event, or risk of disease development, progression or regression. _{42. The method of claim 40 or claim 41, further comprising:} accessing one or more additional risk metrics of the subject; and inputting the one or more additional risk metrics into the trained machine learning or statistical model together with the metric indicative of the number of _{vessels or segments thereof having disease and/or vascular state, preferably} selected from inflammation, oedema, fibrosis, vascularity, lipolysis and/or adipogenesis, to determine the presence or risk of disease and/or vascular state in a subject. _{43. The method of claim 42, wherein the one or more additional risk metrics} _{comprises one or more plaque features of the blood vessel.} _{44. The method of claim 42 or claim 43, wherein the one or more additional} risk metrics further comprises one or more risk factors or characteristics of the subject. _{45. The method of any of claims 40 to 44, wherein the input to the trained} machine learning or statistical model further comprises one or more vessel features of the blood vessels. _{46. The method of any of claims 40 to 45, wherein the machine learning or} statistical model is trained by: generating a reference dataset that comprises: (i) a plurality of reference metrics indicative of the number of b_{lood vessels, or segments thereof having disease or vascular state,} preferably selected from one or more of: inflammation, oedema, fibrosis, v_{ascularity, lipolysis and/or adipogenesis, from two or more blood vessels} _{of the subject; and} _{(ii) a plurality of reference classifications of the presence or risk} _{of disease and/or vascularity, associated with the reference metrics} _{indicative of the number of blood vessels or segments thereof having} disease or vascular state; and training the machine learning or statistical model using the reference dataset to learn a relationship between the metrics indicative of the number of blood vessels _{or segments thereof having disease and/or vascular state, and the classifications} of the presence or risk of disease and/or vascular state. _{47. A method of training a machine learning or statistical model for} determining the presence or risk of disease and/or vascular state, comprising: generating a reference dataset that comprises: (i) a plurality of reference metrics indicative of the number of b_{lood vessels, or segments thereof having disease and/or vascular state,} preferably selected from: inflammation, oedema, fibrosis, vascularity, lipolysis and/or adipogenesis, from two or more blood vessels of the s_{ubject; and} _{(ii) a plurality of reference classifications of the presence or risk} _{of disease and/or vascular state associated with the reference metrics} _{indicative of the number of blood vessels or segments thereof having} disease and/or vascular state; and training the machine learning or statistical model using the reference _{dataset to learn a relationship between the metrics indicative of the number of} _{blood vessels or segments thereof having disease and/or vascular state, and the} classifications of the presence or risk of disease and/or vascular state. _{48. The method of claim 46 or claim 47, wherein the reference dataset further} comprises one or more reference additional risk metrics of the subject. _{49. The method of any of claims 40 to 48, wherein the metric indicative of the} _{number of blood vessels, or segments thereof having disease or vascular state} _{are selected from: inflammation, oedema, fibrosis, vascularity, lipolysis and/or} _{adipogenesis of the subject comprises an inflammation, oedema, fibrosis,} _{vascularity, lipolysis and/or adipogenesis metric of each of two or more blood} _{vessels, or segments thereof, of the subject.} _{50. The method of any of claims 40 to 49, wherein the machine learning or} statistical model is a regression or classifier model, preferably a Cox Proportional- Hazards regression model _{51. A machine learning or statistical model trained using the method of claim} 47. 52. A system for determining the presence or risk of disease and/or vascular state, comprising at least one processor in communication with at least one memory device, the at least one memory device having stored thereon instructions for causing the at least one processor to perform a method according to any of claims 40 to 52.

ABSTRACT METHOD The invention relates to computer implemented methods for determining the presence or risk of disease and/or vascular state in a subject. The method comprises (a) using data gathered from a computer tomography (CT) scan along a length of two or more blood vessels to determine: (i) a value for the radiodensity and/or one or more radiomic features of the perivascular tissue surrounding each blood vessel or segment thereof; (b) using the value for the radiodensity and/or one or more radiomic features of the perivascular tissue surrounding each blood vessel or segment thereof to determine the presence of disease and/or vascular state in each blood vessel, or segment thereof; and (c) generating an output value based on the number of vessels or segments thereof with disease and/or vascular state that indicates the subject’s presence or risk of disease and/or vascular state. _{The invention also relates to computer program products, systems and machine} _{learning or statistical modeals for performing the methods of the invention. Also} disclosed is methods of training said machine learning or statistical models, and related systems.

indicative of the number of blood vessels or segments thereof having disease or vascular state; and training the machine learning or statistical model using the reference dataset to learn a relationship between the metrics indicative of the number of blood vessels or segments thereof having disease and/or vascular state, and the classifications of the presence or risk of disease and/or vascular state.

47. A method of training a machine learning or statistical model for determining the presence or risk of disease and/or vascular state, comprising: generating a reference dataset that comprises:

(i) a plurality of reference metrics indicative of the number of blood vessels, or segments thereof having disease and/or vascular state, preferably selected from: inflammation, oedema, fibrosis, vascularity, lipolysis and/or adipogenesis, from two or more blood vessels of the subject; and

(ii) a plurality of reference classifications of the presence or risk of disease and/or vascular state associated with the reference metrics indicative of the number of blood vessels or segments thereof having disease and/or vascular state; and training the machine learning or statistical model using the reference dataset to learn a relationship between the metrics indicative of the number of blood vessels or segments thereof having disease and/or vascular state, and the classifications of the presence or risk of disease and/or vascular state.

48. The method of claim 46 or claim 47, wherein the reference dataset further comprises one or more reference additional risk metrics of the subject.

49. The method of any of claims 40 to 48, wherein the metric indicative of the number of blood vessels, or segments thereof having disease or vascular state are selected from: inflammation, oedema, fibrosis, vascularity, lipolysis and/or adipogenesis of the subject comprises an inflammation, oedema, fibrosis, vascularity, lipolysis and/or adipogenesis metric of each of two or more blood vessels, or segments thereof, of the subject. 98

50. The method of any of claims 40 to 49, wherein the machine learning or statistical model is a regression or classifier model, preferably a Cox Proportional- Hazards regression model

51 . A machine learning or statistical model trained using the method of claim 47.

52. A system for determining the presence or risk of disease and/or vascular state, comprising at least one processor in communication with at least one memory device, the at least one memory device having stored thereon instructions for causing the at least one processor to perform a method according to any of claims 40 to 52.