WO2025002963A1 - Abnormality detection in medical images - Google Patents

Abstract

A method is provided of abnormality detection in a medical image. Different regions of the image are associated with different abnormalities. Correlations are derived between the presence of the respective associated abnormalities in the different regions of a training dataset. For a target abnormality, a region of interest is defined as a combination of the regions for which a correlation of the respective associated abnormality with the target abnormality meets a threshold. That combined region of interest is then used for detecting the target abnormality in the medical image.

Description

ABNORMALITY DETECTION IN MEDICAL IMAGES

FIELD OF THE INVENTION

This invention relates to the detection of abnormalities in medical images. For example, it relates to the analysis of medical images using AL

BACKGROUND OF THE INVENTION

For Al-based reporting of medical scan images, for example MR knee exams, a large number of neural network (NN) - based classifiers must be trained. Training and performance of the NN- based classifiers often improve when the region-of-interest (ROI) in the image where a finding (e.g.. an anatomical abnormality) to be detected is properly defined (e.g. by a bounding box). These findings usually relate to specific anatomical structures (e.g. meniscus, ligament, cartilage on a specific part of the bone, etc.). The ROI for investigating a finding should thus be restricted to the associated anatomical structure.

A segmentation of the relevant anatomical structures is performed for example using model-based segmentation. The label mask corresponding to a specific anatomical structure and related findings is used to define a ROI that is the input to a classification network for detecting the presence of a specific finding.

Although findings usually refer to specific structures, in various cases a finding in one structure strongly correlates with a finding in another structure. For instance, cartilage damage correlates with bone subchondral edema. For those cases, restricting the region-of-interest to the primary anatomical structure related to the finding can lead to suboptimal classification results.

There is therefore a need for an improved way to detect a target abnormalities (i.e., a target finding) using machine learning.

SUMMARY OF THE INVENTION

The invention is defined by the claims.

According to examples in accordance with an aspect of the invention, there is provided a method of training a classifier to detect a target abnormality in a medical image, wherein the medical image comprises a plurality of regions corresponding to different anatomical features and wherein at least one region of the plurality of regions has at least one associated abnormality, wherein the method comprises: receiving a training dataset; using the training dataset to derive correlations between the presence of the respective associated abnormalities in the different regions; and for the target abnormality, defining a region of interest as a combination of the regions for which a correlation of the respective associated abnormality with the target abnormality meets a threshold; and training the classifier using the training dataset to detect the target abnormality using the region of interest

This method relates to the training of the classifier for subsequent use in detecting the target abnormality in a medical image.

This method exploits correlations between detected abnormalities in different anatomical features in order to define the region of interest used for finding a specific target abnormality.

To derive the correlations, an anatomical segmentation is performed using neural networks, model-based segmentation or other techniques that label the anatomical features (e.g. meniscus, ligament, cartilage on a specific part of the bone, ... ) for which abnormalities should be detected. Correlations between those abnormalities can then be analyzed in the dataset.

When there is a high correlation (above a threshold) of the target abnormality with other abnormalities, the regions are merged (e.g., anatomical label masks are merged) to define the region of interest for detecting the target abnormality. Thus, rather than detecting a target abnormality using the segmented image region for the particular anatomical feature primarily associated with that abnormality, other regions are also used in which other abnormalities correlate with the target abnormality. If there is no correlation with other abnormalities, the region of interest will remain as the single region primarily associated with the target abnormality.

The correlation information may also be used to decide whether independent classifiers are used for different abnormalities (e.g., when there is no or little correlation) or a multi-class classifier should be used allowing to detect multiple findings (e.g., when there is strong correlation between findings).

The trained classifier is then used to detect a target abnormality. In this way, the detection performance for identifying the target abnormality is optimized. The classifier for example comprises a neural network.

The method may further comprise training a second classifier using the training dataset and using the region having the target abnormality as the region of interest of the second classifier.

In this way, classifiers are trained based on a fused region of interest as well as based on a single region of interest of a single anatomical feature. In this way, the user can be presented with multiple classification results which have been derived in different ways.

The invention also provides a computer program comprising computer program code which is adapted, when said program is run on a computer, to implement the method defined above. The invention also provides a method of abnormality detection in a medical image, for detecting a target abnormality in the medical image, wherein the medical image comprises a plurality of regions corresponding to different anatomical features and wherein at least one regions of the plurality of regions has at least one associated abnormality, wherein the method comprises: performing segmentation of anatomical structures and constructing regions of interest on the basis of the segmentation result; and applying the classifier trained using the method defined above to the medical image.

By using the correlations as explained above between detected abnormalities in different anatomical features, the detection performance of the trained neural network-based classifier for identifying the target abnormality is optimized.

The detection method may further comprise applying a second classifier trained using the training dataset and using the region having the target abnormality as the region of interest of the second classifier. Thus, classifiers may be applied that have been trained based on a fused region of interest as well as based on a single region of interest of a single anatomical feature.

The method may then comprise presenting classification probabilities to the user for the classifier and for the second classifier. Also, the classification probabilities could be fused (e.g., averaged) before providing the final classification result.

The method may further comprise representing the region of interest on which the classifier operates on a display. This enables the user to identify the anatomical features which are being analyzed by the classifier in order to detect the target abnormality.

The invention also provides a classifier training apparatus for training a classifier to detect a target abnormality in a medical image, wherein the medical image comprises a plurality of regions corresponding to different anatomical features and wherein at least one region of the plurality of regions has at least one associated abnormality, wherein the apparatus comprises a processor configured to: receive a training dataset; use the dataset to derive correlations between the presence of the respective associated abnormalities in the different regions; for the target abnormality, define a region of interest as a combination of the regions for which a correlation of the respective associated abnormality with the target abnormality meets a threshold; and train the classifier using the training dataset to detect the target abnormality using the region of interest

The invention also provides an image analysis system for analyzing a medical image to detect a target abnormality, wherein the medical image comprises a plurality of regions corresponding to different anatomical features and wherein at least one region of the plurality of regions has at least one associated abnormality, the system comprising a processor configured to: perform segmentation of anatomical structures and construct regions of interest on the basis of the segmentation result; and apply the classifier trained using the method defined above to the medical image.

The trained classifier may for example be a multi-class classifier. This is a classifier that is able simultaneously to provide a response about the presence or absence of more than one finding.

The processor is for example further configured to: apply a second classifier trained using the training dataset and using the region having the target abnormality as the region of interest of the second classifier; and present classification probabilities to the user for the classifier and for the second classifier.

The processor is for example configured to fuse (e.g., average) classification probabilities for the classifier and the second classifier and provide the fused classification probabilities with the classification result.

The system for example comprises a display and a display controller, wherein the display controller is configured to control the display to represent the region of interest on which the classifier operates.

These and other aspects of the invention will be apparent from and elucidated with reference to the embodiment(s) described hereinafter.

BRIEF DESCRIPTION OF THE DRAWINGS

For a better understanding of the invention, and to show more clearly how it may be carried into effect, reference will now be made, by way of example only, to the accompanying drawings, in which:

Fig. 1 shows a method of training a classifier to detect a target abnormality in a medical image;

Fig. 2 shows a set ofN findings Finding 1 to Finding N (i.e., anatomical abnormalities) and their associated regions, ROI 1 to ROI N, and shows correlations between those findings;

Fig. 3 shows a method of abnormality detection in a medical image; and

Fig. 4 shows a classifier training apparatus and image analysis system.

DETAILED DESCRIPTION OF THE EMBODIMENTS

The invention will be described with reference to the Figures.

It should be understood that the detailed description and specific examples, while indicating exemplary embodiments of the apparatus, systems and methods, are intended for purposes of illustration only and are not intended to limit the scope of the invention. These and other features, aspects, and advantages of the apparatus, systems and methods of the present invention will become better understood from the following description, appended claims, and accompanying drawings. It should be understood that the Figures are merely schematic and are not drawn to scale. It should also be understood that the same reference numerals are used throughout the Figures to indicate the same or similar parts.

The invention provides a method of abnormality detection in a medical image. Different regions of the image are associated with different abnormalities. Correlations are derived between the presence of the respective associated abnormalities in the different regions of a training dataset. For a target abnormality, a region of interest is defined as a combination of the regions for which a correlation of the respective associated abnormality with the target abnormality meets a threshold. That combined region of interest is then used for detecting the target abnormality in the medical image.

The term "abnormality" is used to describe any detectable anatomical characteristic which results from a medical condition or injury, and thus distinguishes from a healthy injury-free subject. Each or at least some of such abnormality (-ies) may more generally be considered to be a finding, i.e., a presence of a particular characteristic.

The invention exploits correlations between different abnormalities (i.e., findings) to define a ROI where a specific finding (i.e., a target abnormality) should be detected. The approach is used during training and NN-based inference.

At least in some aspects, the invention provides for a system to exploit correlations between findings to define the Region of Interest, ROI, where a specific finding should be detected. The approach might be used in some embodiments during training and NN-based inference. In particular, an anatomical segmentation may be performed using e.g. neural networks or model-based segmentation that labels the anatomical structures (e.g. meniscus, ligament, cartilage on a specific part of the bone, ...) for which findings should be detected. In addition, correlations between the findings in the training set are analysed. When there is a high correlation (above a threshold) of the considered finding with other findings, respective anatomical label masks are merged to define the ROI for detecting the finding, while this is not done when correlation is low (below a threshold). The correlation information may also be used to decide whether independent classifiers are used for different findings (preferred when there is no or little correlation) or a multi-class classifier allowing to detect multiple findings is used (preferred when there is strong correlation between findings). Hence, the detection performance of the NN-based classifiers for identifying findings is optimized at least in some embodiments.

At least in some aspects, the invention describes a method of training a classifier to detect a target abnormality in a medical image, wherein the medical image comprises a plurality of regions corresponding to different anatomical features and wherein at least one of the plurality of regions has at least one associated abnormality. The method comprises the steps of: receiving (10) a training dataset; using (12) the training dataset to derive correlations between the presence of the respective associated abnormalities in the different regions; for the target abnormality, defining (14) a region of interest as a combination of the regions for which a correlation of the respective associated abnormality with the target abnormality meets a threshold; and training (16) the classifier using the training dataset to detect the target abnormality using the region of interest. In another aspect, a method for detecting a target abnormality in the medical image is disclosed, wherein the medical image comprises a plurality of regions corresponding to different anatomical features and wherein at least one ofthe plurality of regions has at least one associated abnormality. The method comprises the steps of: performing (21) segmentation of anatomical structures in the at least one of the plurality of regions and constructing regions of interest on the basis of the segmentation result; and applying (22) the classifier trained using the method of any one of claims 1-3 to the regions of interest of the medical image. Further the invention will be described referring to single regions, but it is to be understood that it is given for illustration purposes only and the invention is to be understood to be applicable at lease to some of the plurality of regions, at least to one region of the plurality of regions, to all regions of the plurality of regions or any other combination. The following descriptions of regions is not limiting in any way.

Fig. 1 shows a method for training a classifier to detect a target abnormality in a medical image. The medical image comprises a plurality of regions corresponding to different anatomical features and each region has at least one associated abnormality.lt is to be understood that word “each” is not limiting in any way and is provided for illustration purposes only. The application refers generally to at least one region of the plurality of regions.

A training dataset is received in step 10.

The dataset indicates, for different images, which findings (i.e., abnormalities) are present in the various regions of the images.

From this dataset, correlations are derived in step 12 between the presence of the respective associated abnormalities in the different regions.

The correlations could be standard sample correlation coefficients. For binary variables (i.e., finding present or not present) this results in the phi coefficient (or mean square contingency coefficient). Any suitable correlation metric may be used, and the findings are not limited to binary findings. The findings may instead have a value representing a severity of the abnormality.

For a target abnormality, a region of interest is defined in step 14 as a combination of the regions for which a correlation of the respective associated abnormality with the target abnormality meets a threshold.

The classifier is trained in step 16 using the training dataset to detect the target abnormality using the region of interest.

For the detection of abnormalities in the different regions of the training dataset, an anatomical segmentation is performed to differentiate the regions using neural networks, model-based segmentation or other techniques that label the anatomical structures (e.g. meniscus, ligament, cartilage on a specific part of the bone, ... ) for which findings should be detected.

When there is a high correlation between the findings in the training set (above a threshold) of the target abnormality (i.e., target finding) with other abnormalities (i.e., other findings), respective anatomical label masks are merged to define the ROI for detecting the target abnormality. This is not done when the correlation is low, below a threshold.

The correlation information may also be used to decide whether independent classifiers are used for different findings (preferred when there is no or little correlation) or a multi-class classifier allowing to detect multiple findings is used (preferred when there is strong correlation between findings). In that way, detection performance of the NN-based classifiers for identifying findings is optimized.

In some embodiments, correlations between the findings in the training set are analysed. As an example and as indicated in the previous embodiment, when there is a high correlation (above a threshold) of the considered finding with other findings, respective anatomical label masks are merged to define the ROI for detecting the finding, while this is not done when correlation is low (below a threshold). In some embodiments, the correlation information may also be used to decide whether independent classifiers are used for different findings (preferred when there is no or little correlation) or a multiclass classifier allowing to detect multiple findings is used (preferred when there is strong correlation between findings).

At least in some embodiments, the correlation information and ground truth annotation may also be used to set-up the classifiers. When there is no or little correlation between findings, independent classifiers may be preferred. When there is strong correlation between findings or classes / findings are mutually exclusive as indicated, for instance, by the ground truth annotations, multi-class classifiers addressing two (or more) findings might be used. Defining the presence / absence of finding 1 by Tl/ N1 and the presence / absence of finding 2 by T2/ N2, classes for the multi-class classification problem can be defined by {(Nl, N2), (Tl, N2), (Nl, T2), (Tl, T2)} or by {(Nl, N2), (Tl, N2), (Nl, T2)} when the two classes are mutually exclusive.

The use of correlations may be understood from a correlation table as shown in Fig. 2, which illustrates findings derived from the training dataset.

Fig. 2 shows a set ofN findings Finding 1 to Finding N (i.e., anatomical abnormalities) and their associated regions, ROI 1 to ROI N.

The numbers in the table relate to correlation values.

The final ROI for the target abnormality is then defined by the union of all ROIs in the corresponding column or row for which the correlation value is above a threshold.

For the example shown, and a threshold of 0.5, the final ROI for finding x would, for instance, be defined by the union of ROI x and ROI y. The threshold may be lower, such as 0.4 or 0.3 for a simple correlation coefficient. The final ROI itself may them be the bounding box of ROI x and ROI y. Within the combined bounding box, pixels / voxels neither belonging to ROI x or ROI y may be masked.

The final ROI is used together with the ground truth classification results and findings for NN training or together with the trained NN for inferring findings.

There are various ways to combine information from several ROIs available to a NN without merging them directly into a single ROI. For instance, a NN may be structured to use the information of several ROIs separately as input.

The correlation information and ground truth annotation may also be used to set up the classifiers. When there is no or little correlation between findings, independent classifiers may be preferred, used and trained. When there is a strong correlation between findings or classes, or findings are mutually exclusive as indicated for instance by the ground truth annotations, multi-class classifiers addressing two (or more) findings might be used. For example, defining the presence / absence of Finding 1 by Tl/ N1 and the presence / absence of Finding 2 by T2/ N2, classes for the multi-class classification problem can be defined by {(Nl, N2), (Tl, N2), (Nl, T2), (Tl, T2)} or by {(Nl, N2), (Tl, N2), (Nl, T2)} when the two classes are mutually exclusive.

The elements of the correlation table may also used to weigh different terms in the objective/loss function for NN training.

By way of example, typical correlations between findings for routine MR knee examinations are, for instance:

Cartilage thickness loss / defect vs. subchondral edema - cor -0.33.

Meniscus tear vs. Cartilage thickness loss / defect - cor -0.25.

ACL high grade sprain vs. bone fracture / contusion / dislocation - cor - 0.33. Periarticular cyst vs. Cartilage full thickness loss / defect - cor - 0.24.

Instead of using only simple correlation values as described above, it is also possible to estimate and consider conditional probabilities from the correlation data. For example, finding 1 may have a frequency of 0.7, finding 2 may have a frequency of 0.2 and finding 1 is present when finding 2 is present, i.e., cor = 0.33, p(f2|fl=true) = 0.29 and p(fl|f2=true) = 1.

Using conditional probabilities, regions might for example only be merged for the neural network of a finding if the conditional probability given the finding is above a threshold (e.g., 0.5).

Multiple classifiers may be trained for one finding using different combinations of regions derived from the correlation analysis, and the final classification may be obtained by ensembling classifier results.

An overall system may have a large number of neural networks (e.g., 500 or more). For use of the neural network by a customer, the table that has been derived from the correlation matrix can be provided and this indicates which regions should be used for which neural network. For a particular neural network, or a related finding, the actual region being used by the neural network can be displayed to the user.

There are various datasets used in the training and use of the neural network. A first dataset is for deriving the correlations, a second dataset has images and information about the findings, used for training of the neural network-based classifiers. A third dataset is of medical images where the classifier (after training) will be applied.

The first dataset does not need to include images. In addition, the second dataset could be suited to derive the correlations, but it might also only indicate the presence or absence of a distinct (set of) findings and not allow to derive correlations.

The training of a neural network-based classifier may include numerical execution of the classifier. In that case, the medical image of the third set may also be part of the second dataset (which may thus be split into training, validation and test cases).

Fig. 3 shows a method of abnormality detection in a medical image. The medical image again comprises a plurality of regions corresponding to different anatomical features.

The medical image is received in step 20.

In step 21, properly sampled ROIs are constructed from the medical image. Segmentation of anatomical structures is performed and ROIs are then constructed on the basis of the segmentation result. For example, each ROI could be a bounding box around a segmented anatomical structure, possibly with an additional margin. Within such a bounding box, all voxels / pixels that do not belong to the anatomical structure (and possibly a margin) may be “blanked out” (i.e., replaced by 0). The ROI could also be resampled in a pre-defined way to represent the area of the anatomical structure.

In step 22, the classifier trained using the method described above is applied to the combination of the regions of interest as identified by the correlation analysis.

It may additionally be of interest for a user to know the classification result for the conventional classifier, i.e., a second classifier trained using the region having the target abnormality as the region of interest.

Thus, in step 23, a classifier trained using a region of interest corresponding to a single anatomical feature is applied.

Classification probabilities are then be presented to the user in step 24 for the different classifiers, i.e., the classifier using a fused region of interest and a second classifier using a region of interest corresponding to a single anatomical feature. The user can be presented with multiple classification results which have been derived in different ways. The different classifiers may be ensembled in various ways in order to improve performance. The difference or deviation between different results (in terms of probabilities) also assists in gauging the reliability of the results. In some other embodiments, instead or in addition to fusion, the following techniques can be used: image overlay, image co-registration. Medical images can be of any one of or any one of combination of the following modalities: X-ray, Computed Tomography, Magnetic Resonance Imaging, Ultrasound, Pathology, Positron Emission Tomography, Single-photon Emission Computed Tomography, Angiography.

In step 25, the (fused) region of interest on which the classifier operates is represented on a displayed anatomical image. This enables the user to identify the different anatomical features which have been taken into account in arriving at the classification result.

Fig. 4 shows a training apparatus and image analysis system. As a non-limiting example, this embodiment shows an example of a classifier training apparatus. But other types of training apparatuses could be used within the context of the present application.

The classifier training apparatus comprises a training processor 40 for training a classifier based on a received training dataset 42.

The image analysis system comprises an image analysis processor 44 for analyzing a medical image 46 to detect a target abnormality

The invention relates generally to the use of machine learning. Neural networks are given as an example above, but decision trees or other classifiers may be used.

A machine-learning algorithm is any self-training algorithm that processes input data in order to produce or predict output data. Here, the input data comprises medical image data and the output data comprises detection of a target abnormality in the medical image.

Suitable machine-learning algorithms for being employed in the present invention will be apparent to the skilled person. Examples of suitable machine-learning algorithms include decision tree algorithms and artificial neural networks. Other machine-learning algorithms such as logistic regression, support vector machines or Naive Bayesian models are suitable alternatives.

The structure of an artificial neural network (or, simply, neural network) is inspired by the human brain. Neural networks are comprised of layers, each layer comprising a plurality of neurons. Each neuron comprises a mathematical operation. In particular, each neuron may comprise a different weighted combination of a single type of transformation (e.g. the same type of transformation, sigmoid etc. but with different weightings). In the process of processing input data, the mathematical operation of each neuron is performed on the input data to produce a numerical output, and the outputs of each layer in the neural network are fed into the next layer sequentially. The final layer provides the output.

Methods of training a machine-learning algorithm are well known. Typically, such methods comprise obtaining a training dataset, comprising training input data entries and corresponding training output data entries. An initialized machine-learning algorithm is applied to each input data entry to generate predicted output data entries. An error between the predicted output data entries and corresponding training output data entries is used to modify the machine-learning algorithm. This process can be repeated until the error converges, and the predicted output data entries are sufficiently similar (e.g. ±1%) to the training output data entries. This is commonly known as a supervised learning technique. For example, where the machine-learning algorithm is formed from a neural network, (weightings of) the mathematical operation of each neuron may be modified until the error converges. Known methods of modifying a neural network include gradient descent, backpropagation algorithms and so on.

Variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing the claimed invention, from a study of the drawings, the disclosure and the appended claims. In the claims, the word "comprising" does not exclude other elements or steps, and the indefinite article "a" or "an" does not exclude a plurality.

Functions implemented by a processor may be implemented by a single processor or by multiple separate processing units which may together be considered to constitute a "processor". Such processing units may in some cases be remote from each other and communicate with each other in a wired or wireless manner.

The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage.

A computer program may be stored/distributed on a suitable medium, such as an optical storage medium or a solid-state medium supplied together with or as part of other hardware, but may also be distributed in other forms, such as via the Internet or other wired or wireless telecommunication systems.

If the term "adapted to" is used in the claims or description, it is noted the term "adapted to" is intended to be equivalent to the term "configured to". If the term "arrangement" is used in the claims or description, it is noted the term "arrangement" is intended to be equivalent to the term "system", and vice versa.

In a networked deployment, the computer system operates in the capacity of a server, or as a client user computer in a server-client user network environment, or as a peer computer system in a peer-to-peer or distributed network environment. The computer system can also be implemented as or incorporated into various devices, such as a server or another type of computer such as a workstation that includes a controller, a stationary computer, a mobile computer, a personal computer (PC), a laptop computer, a tablet computer, or any other machine capable of executing a set of software instructions sequentially or non-sequentially that specify actions to be taken by that machine. The computer system can be incorporated as an integrated system part of a larger system that includes additional devices. In an embodiment, the computer system can be implemented using electronic devices that provide voice, video, or data communication possibilities. Further, while the computer system is illustrated in the singular, the term “system” shall also be taken to include any collection of systems or sub-systems that individually or jointly execute a set or multiple sets, of software instructions to perform one or more computer functions.

The computer system may also include a processor. The processor executes instructions to implement some, or all aspects of methods and processes described herein. The processor is tangible and non-transitory. As used herein, the term “non-transitory” is to be interpreted not as an eternal characteristic of a state, but as a characteristic of a state that will last for a period. The term “non- transitory” specifically disavows fleeting characteristics such as characteristics of a carrier wave or signal or other forms that exist only transitorily in any place at any time. The processor is an article of manufacture and/or a machine component. The processor is configured to execute software instructions to perform functions as described in the various embodiments herein. The processor may be a general- purpose processor or may be part of an application specific integrated circuit (ASIC). The processor may also be a microprocessor, a microcomputer, a processor chip, a controller, a microcontroller, a digital signal processor (DSP), a state machine, or a programmable logic device, a logical circuit, including a programmable gate array (PGA), such as a field programmable gate array (FPGA), or another type of circuit that includes discrete gate and/or transistor logic. The processor may be a central processing unit (CPU), a graphics processing unit (GPU), or both. Additionally, any processor described herein may include multiple processors, parallel processors, or both. Multiple processors may be included in, or coupled to, a single device or multiple devices. The processor can include one or more internal levels of cache, and a bus controller or bus interface unit to direct interaction with a bus. The term “processor” as used herein encompasses an electronic component able to execute a program or machine executable instruction. References to a computing device comprising “a processor” should be interpreted to include more than one processor or processing core, as in a multi -core processor. A processor may also refer to a collection of processors within a single computer system or distributed among multiple computer systems. The term computing device should also be interpreted to include a collection, or network, of computing devices each including a processor or processors. Programs have software instructions performed by one or multiple processors that may be within the same computing device or which may be distributed across multiple computing devices. Further, the software instructions, when executed by the processor, perform one or more steps of the methods and processes as described herein.

The computer system further includes a main memory and a static memory, where memories in the computer system communicate with each other and the processor via a bus. Either or both main memory and the static memory may be considered representative examples of the memory of the controller, and store instructions used to implement some, or all aspects of methods and processes described herein. Memories described herein are tangible storage mediums for storing data and executable software instructions and are non-transitory during the time software instructions are stored therein. The main memory and the static memory are articles of manufacture and/or machine components. The main memory and the static memory are computer-readable mediums from which data and executable software instructions can be read by a computer (or e.g., the processor). Each of the main memory and the static memory may be implemented as one or more of random access memory (RAM), read only memory (ROM), flash memory, electrically programmable read only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, a hard disk, a removable disk, tape, compact disk read only memory (CD-ROM (Compact Disk - Read Only Memory)), digital versatile disk (DVD), floppy disk, Blu-ray disk, or any other form of storage medium known in the art. The memories may be volatile or non-volatile, secure and/or encrypted, unsecure and/or unencrypted.

The illustrations of the embodiments described herein are intended to provide a general understanding of the structure of the various embodiments. The illustrations are not intended to fully describe all the elements and features of the disclosure described herein. Many other embodiments may be apparent to those of skill in the art upon reviewing the disclosure. Other embodiments may be utilized and derived from the disclosure, such that structural and logical substitutions and changes may be made without departing from the scope of the disclosure. Additionally, the illustrations are merely representational and may not be drawn to scale. Certain proportions within the illustrations may be exaggerated, while other proportions may be minimized. Accordingly, the disclosure and the figures are to be regarded as illustrative rather than restrictive.

Although specific embodiments have been illustrated and described herein, it should be appreciated that any subsequent arrangement designed to achieve the same or similar purpose may be substituted for the specific embodiments shown. This disclosure is intended to cover all subsequent adaptations or variations of various embodiments. Combinations of the above embodiments, and other embodiments not specifically described herein, will be apparent to those of skill in the art.

The preceding description of the disclosed embodiments is provided to enable any person skilled in the art to practice the concepts described in the present disclosure. As such, the above disclosed subject matter is to be considered illustrative, and not restrictive, and the appended claims are intended to cover all such modifications, enhancements, and other embodiments which fall within the scope of the present disclosure. Thus, to the maximum extent allowed by law, the scope of the present disclosure is to be determined by the broadest permissible interpretation of the following claims and their equivalents and shall not be restricted or limited by the foregoing detailed description.

Any reference signs in the claims should not be construed as limiting the scope.

Claims

CLAIMS:

1. A method of training a classifier to detect a target abnormality in a medical image, wherein the medical image comprises a plurality of regions corresponding to different anatomical features and wherein at least one of the plurality of regions has at least one associated abnormality, wherein the method comprises: receiving (10) a training dataset; using (12) the training dataset to derive correlations between the presence of the respective associated abnormalities in the different regions; for the target abnormality, defining (14) a region of interest as a combination of the regions for which a correlation of the respective associated abnormality with the target abnormality meets a threshold; and training (16) the classifier using the training dataset to detect the target abnormality using the region of interest.

2. The method of claim 1, wherein the classifier comprises a neural network.

3. The method of claim 1 or 2, further comprising training a second classifier using the training dataset and using the region having the target abnormality as the region of interest of the second classifier.

4. A computer program comprising computer program code which is adapted, when said program is run on a computer, to implement the method of any one of claims 1 to 3.

5. A method, for detecting a target abnormality in the medical image, wherein the medical image comprises a plurality of regions corresponding to different anatomical features and wherein at least one ofthe plurality of regions has at least one associated abnormality, wherein the method comprises: performing (21) segmentation of anatomical structures in the at least one of the plurality of regions and constructing regions of interest on the basis of the segmentation result; and applying (22) the classifier trained using the method of any one of claims 1-3 to the regions of interest of the medical image.

6. The method of claim 5, further comprising applying a second classifier trained using a training dataset and using the at least one of the plurality of regions having the target abnormality of the second classifier.

7. The method of claim 6, comprising presenting classification probabilities to the user for the classifier and for the second classifier.

8. The method of any one of claims 5 to 7, representing the region of interest on which the classifier operates on.

9. A training apparatus for training a classifier to detect a target abnormality in a medical image, wherein the medical image comprises a plurality of regions corresponding to different anatomical features and wherein at least one of the plurality of regions has at least one associated abnormality, wherein the apparatus comprises a processor configured to: receive a training dataset; use the training dataset to derive correlations between the presence of the respective associated abnormalities in the different regions; and for the target abnormality, define a region of interest as a combination of the regions for which a correlation of the respective associated abnormality with the target abnormality meets a threshold; and train the classifier using the training dataset to detect the target abnormality using the region of interest.

10. The apparatus of claim 9, wherein the processor is further configured to train a second classifier using the training dataset and using the region having the target abnormality as the region of interest of the second classifier.

11. An image analysis system for analyzing a medical image, wherein the medical image comprises a plurality of regions corresponding to different anatomical features and wherein at least one of the plurality of regions have at least one associated abnormality, the system comprising a processor configured to: perform (21) segmentation of anatomical structures in the at least one of the plurality of regions and construct regions of interest on the basis of the segmentation result; and apply (22) the classifier trained using the method of any one of claims 1-3 to the regions of interest of the medical image.

12. The system of claim 11, wherein the trained classifier is a multi-class classifier with, optionally, multioutput multilabel classification.

13. The system of any one of claims 11 or 12, wherein the processor is further configured to: apply a second classifier trained using the training dataset and using at least one of the plurality of the regions having the target abnormality as the region (-s) of interest of the second classifier; and present classification probabilities to the user for the classifier and for the second classifier.

14. The system of claim 13, wherein the processor is configured to fuse or otherwise combine classification probabilities for the classifier and the second classifier and provide the fused classification probabilities with the classification result.

15. The system of any one of claims 11 to 14, comprising a display and a display controller, wherein the display controller is configured to control the display to represent the at least one of the plurality of regions of interest on which the classifier operates.