RU2847440C2 - Methods and systems for reducing artifacts in image reconstruction - Google Patents

Abstract

FIELD: image reconstruction.

SUBSTANCE: the claimed solution provides for a detector assembly, the sensitive elements of which are designed to detect photons associated with an object, create an initial reconstruction of a tomographic image based on response lines (LoRs), wherein the tomographic image includes a plurality of voxels associated with a volume or region analysed by the detector assembly; determining a system matrix that correlates each voxel in the tomographic image with the sensitive elements; create an attenuation map, wherein the attenuation map is based on the tomographic image and is defined as a set of voxels, wherein each voxel of the attenuation map contains information about the voxel relating to the degree or probability of attenuation and/or scattering associated with that voxel; perform an iterative tomography reconstruction algorithm to update the tomography image, wherein the iteration includes updating a system matrix based on the attenuation map voxel information; update the attenuation map based on the updated tomography image.

EFFECT: reduction of artefacts in PET image reconstruction.

31 cl, 11 dwg

Description

AREA OF TECHNOLOGY

The present invention generally relates to image reconstruction, including positron emission tomography (PET) and computed tomography (CT). More specifically, the present invention relates to a method for reducing artifacts in tomographic image reconstruction. The present invention also relates to a system for reducing artifacts in tomographic image reconstruction.

PRIOR ART

One of the challenges in creating tomographic images of objects (using ionizing radiation) is that various physical effects, particularly scattering and attenuation of radiation imaging, lead to distortions or "artifacts" in the reconstructed images, which degrade image quality. Scattering refers to a physical process that changes the direction of particle motion, while attenuation refers to a physical process that reduces the intensity of radiation, for example, by absorbing certain particles.

The two most important types of imaging are positron emission tomography (PET) and computed tomography (CT), hereafter referred to as CT. PET uses radiation generated within an object as a result of nuclear decay, and the scattering and attenuation of the emitted radiation negatively impacts image quality. CT scanning uses X-rays generated on one side of the object and imaged on the other side using a suitable detector system. In this case, X-ray attenuation is a physical process that is measured and therefore does not require compensation, but scattering still results in image degradation.

In this invention, reference is made to PET in a mining environment (MinPET) to explain the claimed invention of attenuation and scatter correction during image reconstruction, as this was the original goal of the present inventors. However, as will be apparent to those skilled in the art, the inventors also recognize that the concepts of the present invention are appropriately applicable to other tomographic processes, such as PET and CT scanning in medicine (clinical and scientific research), since the resulting tomographic images associated with these tomographic processes are also negatively affected by artifacts present therein.

To explain, standard diamond mining operations require enormous amounts of resources, such as water and energy, to process mostly unproductive rock to extract diamonds. Rock processing typically involves a damaging sequence of operations to crush the rock and extract the diamonds, often with relatively low efficiency, for example, approximately 1 carat per ton of processed rock. However, conventional rock crushing can lead to the destruction of diamonds, thereby reducing the profitability of a diamond mine.

Sensor- or detector-based technologies aim to mitigate these unwanted impacts by enabling the early detection of relatively unprocessed diamond-bearing rocks (e.g. kimberlite), which can then be separated and processed in a more environmentally friendly manner that preserves the integrity of the diamond.

Positron emission tomography (PET) can detect minerals, particularly diamonds, in rock. This approach involves irradiating the rock with a beam of gamma rays from bremsstrahlung, such as 40 MeV electrons, or another photon source other than bremsstrahlung, such as inverse Compton scattering, or other technologies. The rock returns to moderate levels of specific radioactivity within a few minutes, by which time the dominant residual activity is PET isotopes. If, after approximately 30 minutes, the PET isotope ^11C exhibits dominant activity, the rock is tested using a detector array to determine the presence or absence of diamond. The detector array typically includes multiple individual sensing elements. For example, a detector array might include a pair of spaced detector arrays arranged parallel to one another, with each array consisting of several individual sensing elements.

The classification data obtained using the PET approach are typically associated with photons detected by a detector array. These photons are emitted from the rock as a result of positron annihilation in the rock. When a PET isotope in the rock emits a positron particle through beta decay, the positron annihilates with a neighboring electron after a path that may involve several scatterings. The most common result of such annihilation is the formation of nearly collinear, oppositely oriented 511 eV gamma-ray photons. Each photon passes through the surrounding material, sometimes changing energy and direction along the way. When the photons reach the detector array, the detector array can output classification data in the form of collision/impact event data with the sensing element.

Typically, two oppositely oriented photons must be detected at approximately the same time on opposite planes of a detector array (e.g., above and below kimberlite being analyzed for diamonds) to form a line of response (LoR). As used herein, the term "event" refers to a photon detected by a single sensing element, where the information may include the energy, position, and time associated with the event. A line of response represents coincident events from opposite detector units that arrived within a short time interval (known as the coincidence window) and are thus considered to originate from the same positron annihilation event. The coincidence window defines the maximum time between coincident events for a given detection system to classify such events as a line of response. In other words, if the difference in the detection times of two collinear events is less than the coincidence window, they are classified as coincident and define a line of response.

In many cases, it is necessary to create/reconstruct a type of image (here referred to as a reconstructed image) from the raw data acquired by the detector system, preferably highlighting areas of interest within such an image. When detecting diamonds using PET, it is desirable to create an image showing the density distribution of the PET isotope within the rock to identify likely diamond-bearing zones.

To reconstruct the PET isotope density distribution image, a three-dimensional (3D) model can be obtained from a four-dimensional (4D) sinogram. This sinogram consists of the complete set of detection data acquired by the detector setup, typically after sorting the response lines by angle and position. For example, a sinogram can be formed by combining the response lines into a set of transverse (x, y) projections for a set of angles (θ, ϕ). In the case where the detector setup contains a pair of detector arrays, as described above, the sinogram is a 4D object specified by two angles in spherical coordinates, θ and ϕ, and two coordinates of the location on the plane midway between the two detector arrays, u and ν. In the context of this description, the 3D model or formed image is called a "reconstructed PET image", "tomographic reconstructed image", "reconstructed tomographic image", "reconstructed image", "tomographic image reconstruction", or "tomographic image".

In the context of PET mineral detection, a reconstructed PET image thus refers to a 3D model of the PET isotope density distribution obtained from a 4D sinogram. In this description, a "projection" refers to a variation of the 4D sinogram corresponding to a specific choice of θ and ϕ angles. A projection is thus a two-dimensional (2D) object specified by u and ν. Furthermore, a "projection element" refers to a single projection element, i.e., a projection value for given u and ν. It represents the cumulative value of all response lines with identical angles and positions, which have been grouped into a single set.

The authors are aware of several technologies that can be used to convert a sinogram into a reconstructed PET image, representing a quantitative 3D representation of source points. They found that the most advanced technology currently available is based on an iterative algorithm designed to find the most probable initial density of 3D source points. The reconstructed PET image is essentially a processed form of the raw data obtained by the detector system, with some information being lost during processing.

As mentioned above, a problem associated with PET image reconstruction using iterative methods is that systematic artifacts are introduced into the reconstructed PET image if the attenuation and scattering effects of two oppositely oriented, collinear, and coincident photons as they pass through the studied medium (e.g., kimberlite) are not taken into account during the reconstruction process. The present authors found that, as a result of these effects, only 3% of the photons can typically be used in the response line construction process. Many attenuated photons are lost, and scattered photons can lead to the construction of a false response line. If the average response line path length in the rock is longer than the X-ray attenuation length, the effect is greater. The path length is affected by gaps in the rock layer and the position of the PET isotope source point in the rock. The authors of the present invention have found that a PET image reconstruction method that does not account for this phenomenon interprets increased expression of response lines associated with open areas or boundary areas as increased concentrations of PET source points. For example, a single spherical rock sample with a uniform distribution of PET isotope source points will be reconstructed as having an increased radial distribution of PET isotope source points.

To demonstrate the problems that can arise when attenuation and/or scattering are not taken into account, consider a pair of sensors, for example, one sensor from a first array in a detector array and the opposite sensor from a second array in a detector array. The number of response lines detected by this pair is proportional to the number of PET isotope nuclei in the line between them. The attenuation or scattering of photons, which reduces the number of detected true response lines, is an exponential decay based on the path length of the photons through the rock. This effect is significantly more pronounced when using PET for diamond detection than when using PET for medical purposes, since rock is much more likely to scatter or absorb photons per unit length than biological tissue. In general, one would expect the number of response lines, N, between sensors to be approximately proportional to the product of the path length, x, since it determines the number of source nuclei, and the exponential decay function from the linear attenuation:

Formula (1) has a maximum at x=1/μ. After this point, adding additional material reduces the signal observed in each pixel of each projection.

However, it should be understood that the real situation is much more complex than the simple formula (1) above implies. In practice, photons are not simply absorbed or not absorbed. Instead, they can travel along complex paths through the rock, causing multiple scattering events, each of which can alter their direction and energy. Therefore, to properly reconstruct images, the effects of attenuation and scattering must be taken into account.

To further illustrate the negative consequences of ignoring the fact of photon attenuation and/or scattering, Figures 1 and 2 show the first example of artifacts that can appear in a reconstructed PET image. Kimberlite sphere 2 with a uniform distribution of diamond-bearing material and a central cavity, shown in Figure 1, was simulated using Geant4's Monte Carlo particle tracking code. This simulation replicates PET isotope annihilations, thermalization, and positron annihilation, followed by tracking the annihilation photons through all materials, including all photon-material interactions, and then ultimately the detector readings. Single events are matched, forming coincident events that form response lines, and they are reconstructed into a PET image using an iterative algorithm that has no attenuation correction. The reconstructed image is shown in Figure 2.

In Fig. 2, which is a "slice" of reconstructed sphere 3 (created using the iterative technique), two types of artifacts are visible. First, the kimberlite near boundaries 4 of sphere 3 is incorrectly enhanced. This is due to the fact that response lines from points near boundaries 4 have a large set of available angles through which they can reach opposite detectors, while the path length through the kimberlite is short. On the other hand, source positions near the center 5 should pass through a large part of the rock diameter, regardless of the direction of the response line. The second type of artifact is located inside a cavity. This cavity is effectively filled with air, and no PET events can occur within it. However, in the reconstructed image of sphere 3, the PET isotope density is shown to be higher in the cavity than outside sphere 3. This is due to the effect of the cavity, which reduces the path length through the kimberlite of the response line that passes through it.

Figures 3 and 4 show a second example of artifacts that can arise when using known iterative reconstruction techniques. Four spheres 6 are arranged tetrahedronally, as can be clearly seen in Figure 3. The center 7 between the spheres 6 should be a region with no activity, since there is no rock there, but when reconstructed, it again shows increased activity, as shown in Figure 4. In other words, the region 9 between the reconstructed spheres 8 should actually show no activity, but as a result of the above-mentioned influences, it is reconstructed incorrectly.

Figures 1-4 are simply examples of the problems caused by the lack of compensation for these effects. With more complex kimberlite arrangements and shapes, image artifacts resulting from attenuation and scattering are significantly more complex.

It is desirable to achieve the highest possible reproduction quality when creating a reconstructed PET image, as this image should represent as accurately as possible the actual PET isotope density distribution in the rock. Therefore, embodiments of the present invention aim to reduce artifacts in PET image reconstruction.

In the context of this description, the term "object" can refer to rock particles, such as kimberlite, regardless of size, or to free diamonds. Therefore, the terms "object," "rock," and "kimberlite" may be used interchangeably in this document. However, the term "object" can also encompass other objects that are visualized or analyzed during the detection method and is therefore not limited to rock. In this regard, an object may include organic tissue, a human or animal body, or a part thereof.

ESSENCE OF THE INVENTION

According to one embodiment of the present invention, a method is provided for reducing artifacts in the reconstruction of tomographic images, wherein the method is carried out in whole or in part by a computer or an automated system connected in communication with a detector installation that contains a plurality of sensitive elements, wherein the sensitive elements are configured to detect photons associated with an object, and the method comprises the following steps:

creating or obtaining an initial reconstruction of a tomographic image associated with an object, wherein the tomographic image is at least partially based on lines of response (LoRs) detected by a detector system, wherein the tomographic image includes a plurality of voxels associated with a volume or region analyzed by the detector system;

determining or obtaining a system matrix that relates each voxel in a tomographic image to sensitive elements;

creating or obtaining an attenuation map, wherein the attenuation map is based on a tomographic image and is defined as a plurality of voxels, wherein each voxel of the attenuation map contains information about the voxel concerning the degree or probability of attenuation and/or scattering associated with that voxel;

performing one iteration of the iterative tomographic reconstruction algorithm to update the tomographic image, wherein performing the iteration includes updating the system matrix based on information about the voxel of the attenuation map;

update the attenuation map based on the updated tomographic image; and

repeating one iteration of the reconstruction algorithm and updating the attenuation map until the image converges to define a tomographic reconstructed image, thereby reducing artifacts in the tomographic image by compensating for or taking into account the attenuation and/or scattering of photons associated with the object.

This method can be used to reduce artifacts in positron emission tomography (PET) image reconstruction by compensating for or accounting for the attenuation and/or scattering of photons emitted from an object. In this regard, the tomographic image is represented as a PET image associated with the object, with the sensing elements configured to detect photons emitted from the object as a result of positron annihilation caused by irradiation of the object with photons of a predetermined energy. The PET image provides a representation of the density distribution of the PET isotope within the volume or region analyzed by the detector system.

The system matrix can associate each voxel in the PET image with every possible pair of sensitive elements.

This method can be used to reduce artifacts in computed tomography (CT) image reconstruction by compensating for or accounting for the scattering of photons from a source as it interacts with an object. In this regard, the tomographic image can be represented as a CT image associated with the object, with the sensing elements configured to detect photons from the source passing through the object. The CT image provides information about the material density and/or electron density in the object analyzed by the detector system. The system matrix can associate each voxel in the CT image with each possible response line from the source to the sensing elements.

In the case of CT, those skilled in the art understand that the response line can be between the X-ray source and the collision position with the detector. In other words, the path of X-rays or X-ray photons from the X-ray source to the detector can constitute the response line.

With CT, the response lines are grouped to produce sinograms, which are reconstructed to produce quantitative 3D images in a manner similar to PET. In this regard, the term "tomographic image," as described here, can thus refer to both PET and CT images. In particular, as described above with respect to PET images, CT images may also be referred to as "reconstructed CT image," "tomographic reconstructed image," "reconstructed tomographic image," or simply "reconstructed image."

It should be understood that a CT image can be a voxelized 3D image, similar to the PET image described here, and therefore the explanations regarding PET images apply accordingly to CT images. The quantity imaged in each voxel can be analogous to the electron density up to first order. It follows that, in the simplest case, a CT image can be a function of the material density and the electron density per element within the material. CT images can then be segmented/divided into different materials. In PET images, the density of PET isotopes within the materials themselves is segmented.

It should be understood that the 3D image itself from a CT scan can be the same or generally similar to an attenuation map. This is because the visualization represents the "attenuation intensity" of materials. When using the present invention for mineral deposit mining, herein referred to as MinPET, both kimberlite and diamond are PET-active, so the PET image/map can be considered an approximate representation of material distributions and, therefore, an attenuation map. When using PET in medicine, the primary PET map is composed of function (how PET isotopes are metabolized in tissue) rather than shape (the material and its shapes). Therefore, this method may include acquiring a CT image to obtain an attenuation map, which is used for correction.

For this reason, PET and CT are often used together in medicine to obtain function and shape for comparison. When correcting a CT image, it represents its own attenuation map, so it is used to correct for the effects of scattering. This method may involve creating or acquiring a CT image of the object and using the CT image as the attenuation map.

When using PET for mining or medicine, both attenuation and scatter are corrected, not just scatter as is the case with CT.

In accordance with another aspect of the present invention, a system is provided for reducing artifacts in the reconstruction of tomographic images, wherein the system comprises a computer or an automated system communicatively connected to a detector installation that comprises a plurality of sensitive elements, wherein the sensitive elements are configured to detect photons associated with an object, wherein the computer or automated system is configured to:

creating or obtaining an initial reconstruction of a tomographic image associated with an object, wherein the tomographic image is at least partially based on lines of response (LoRs) detected by a detector system, wherein the tomographic image includes a plurality of voxels associated with a volume or region analyzed by the detector system;

determining or obtaining a system matrix that relates each voxel in the tomographic image to the sensitive elements;

creating or obtaining an attenuation map, wherein the attenuation map is based on a tomographic image and is defined as a plurality of voxels, wherein each voxel of the attenuation map contains information about the voxel concerning the degree or probability of attenuation and/or scattering associated with that voxel;

performing one iteration of the iterative tomographic reconstruction algorithm to update the tomographic image, wherein performing the iteration includes updating the system matrix based on information about the voxel of the attenuation map;

updating the attenuation map based on the updated tomographic image; and

repeating one iteration of the reconstruction algorithm and updating the attenuation map until the image converges to determine a tomographic reconstructed image, thereby reducing artifacts in the tomographic image by compensating for or taking into account the attenuation and/or scattering of photons associated with the object.

In accordance with another aspect of the present invention, a method is provided for reducing artifacts in the reconstruction of tomographic images, wherein the method is performed by a computer or an automated system communicatively connected to a detector installation that contains a plurality of sensitive elements, wherein the sensitive elements are configured to detect photons associated with an object, wherein the method includes using a machine learning-trained visualization device to create a reconstructed tomographic image associated with the object, wherein the visualization device is trained on at least computer-simulated images and/or object data obtained from a computerized modeling device to compensate for or take into account the attenuation and/or scattering of photons associated with the object, thereby reducing artifacts in the reconstructed image.

According to another aspect of the present invention, a system is provided for reducing artifacts in the reconstruction of tomographic images, wherein the system comprises a computer or an automated system communicatively connected to a detector installation that comprises a plurality of sensing elements, wherein the sensing elements are configured to detect photons associated with an object, wherein the computer or automated system is configured to detect photons associated with an object, wherein the method includes using a machine learning-trained imaging device to create a reconstructed tomographic image associated with an object, wherein the imaging device is trained on at least computer-simulated images and/or object data obtained from a computerized simulator to compensate for or take into account attenuation and/or scattering of photons associated with the object, thereby reducing artifacts in the reconstructed image.

In accordance with another aspect of the present invention, a method is provided for reducing artifacts in the reconstruction of tomographic images, wherein the reconstructed tomographic image is created by a computer or an automated system communicatively connected to a detector system that contains a plurality of sensitive elements, wherein the sensitive elements are configured to detect photons associated with objects, the method including filling cavities between or in the objects with a predetermined material or materials before analyzing the objects with the detector system.

In accordance with another aspect of the present invention, a method is provided for reducing artifacts in the reconstruction of positron emission tomography (PET) images, said method being carried out by a computer or an automated system connected in communication with a detector system that contains a plurality of sensitive elements, wherein the sensitive elements are configured to detect photons emitted from an object as a result of positron annihilation caused by irradiation of the object with photons of a predetermined energy, said method comprising the following steps:

creating or obtaining an initial reconstruction of a PET image associated with an object, wherein the PET image is at least partially based on lines of response (LoRs) detected by a detector system, wherein the PET image includes a plurality of voxels and provides a representation of the density distribution of the PET isotope in a volume or region analyzed by the detector system;

determine or obtain a system matrix that associates each voxel in the PET image with each possible pair of sensitive elements;

creating or obtaining an attenuation map, wherein the attenuation map is based on a PET image and is defined as a plurality of voxels, wherein each voxel of the attenuation map contains information about the voxel concerning the degree or probability of attenuation and/or scattering associated with that voxel;

performing one iteration of the iterative PET reconstruction algorithm to update the PET image, wherein performing the iteration includes updating the system matrix based on the attenuation map voxel information;

updating the attenuation map based on the updated PET image; and repeating one iteration of the reconstruction algorithm and updating the attenuation map until the PET image converges to determine the reconstructed PET image, thereby reducing artifacts in the PET image by compensating for or taking into account the attenuation and/or scattering of photons emitted from the object.

This method may include using additional data from the sensor to create, retrieve, and/or update the attenuation map.

According to another aspect of the present invention, there is provided a method for reducing artifacts in the reconstruction of PET images, wherein the method is performed by a computer or an automated system communicatively connected to a detector setup that contains a plurality of sensing elements, wherein the sensing elements are configured to detect photons emitted from an object as a result of positron annihilation caused by irradiating the object with photons of a predetermined energy, wherein the method includes using a machine learning-trained imaging device to create a reconstructed PET image that provides a representation of the density distribution of a PET isotope in a volume or region analyzed by the detector setup, wherein the method includes using a machine learning-trained imaging device to create a reconstructed tomographic image associated with the object, wherein the imaging device is trained on at least computer-simulated images and/or object data obtained from a computerized simulator to compensate for or take into account the attenuation and/or scattering of photons emitted from the object, thereby reducing artifacts in the reconstructed PET image.

According to another aspect of the present invention, a system for reducing artifacts in the reconstruction of PET images is proposed, wherein the system comprises at least one computer or automated system communicatively connected to a detector installation that comprises a plurality of sensitive elements, wherein the sensitive elements are configured to detect photons emitted from an object as a result of positron annihilation caused by irradiation of the object with photons of a predetermined energy, and wherein the computer or automated system is configured to implement, essentially, the method described above.

According to a further aspect of the present invention, there is provided a diamond mine data processing system that includes substantially the system described above.

In accordance with yet another aspect of the present invention, a computer program product is provided, comprising at least one machine-readable medium on which is stored at least one computer program, which, when executed by a computer or automated system, the computer or automated system essentially performs the method as described above. The machine-readable medium may be a non-volatile machine-readable medium.

In accordance with yet another additional aspect of the present invention, a method is provided for reducing artifacts in the reconstruction of PET images, wherein the reconstructed PET image is created by a computer or an automated system communicatively connected to a detector system that contains a plurality of sensitive elements, wherein the sensitive elements are configured to detect photons emitted from objects as a result of positron annihilation caused by irradiation of the object with photons of a predetermined energy, this method including filling cavities between or in the objects with a predetermined material or materials before analyzing the objects with the detector system.

Glass beads, liquid, etc. can be used as materials.

It will be apparent to those skilled in the art that the above comments relating to one aspect of the invention may be applied, correspondingly, to other aspects of the present invention.

BRIEF DESCRIPTION OF DRAWINGS

Fig. 1 shows a three-dimensional image of a modeled kimberlite ball with a central cavity;

Fig. 2 shows a three-dimensional image of a "slice" taken through the reconstructed PET image of the ball shown in Fig. 1, in which the image was reconstructed using a known iterative technique;

Fig. 3 shows a three-dimensional image of four simulated kimberlite balls arranged tetrahedrally;

Fig. 4 shows a three-dimensional image of a "slice" taken through a reconstructed PET image of the composition shown in Fig. 3, in which the image was reconstructed using a known iterative technique;

Fig. 5 is a diagram of an example of a diamond mine data processing system in which embodiments of the present invention may be implemented;

Fig. 6 shows an image of kimberlite on a moving belt between detector arrays, or planes, of a detector setup in accordance with an exemplary embodiment of the invention, showing two coinciding, oppositely oriented 511 eV gamma rays which together form a response line;

Fig. 7 is a diagram illustrating an example of a method that can be used to reduce artifacts in PET image reconstruction by attenuation correction;

Fig. 8 shows examples of segmentation algorithms that may be used in embodiments of the present invention to assist in the reconstruction of PET images;

Fig. 9 shows a schematic diagram of the design concept of a method according to which voxels can be counted in an attenuation map in accordance with embodiments of the invention;

Fig. 10 shows a schematic diagram of a design concept of a method according to which the path length can be determined by detecting an edge in accordance with embodiments of the invention; and

Fig. 10 is a schematic representation of a machine in the exemplary form of a computer system in which a set of instructions may be executed that cause the machine to implement any one or more of the techniques discussed herein.

DETAILED DESCRIPTION WITH LINKS TO DRAWINGS

The following description of the invention is provided for a more complete understanding of the invention. Those skilled in the art will appreciate that many modifications can be made to the described embodiments while still achieving the beneficial effects of the present invention. It will also be apparent that some of the desired advantages of the present invention can be achieved by selecting certain features of the present invention without using other features. Accordingly, those skilled in the art will appreciate that modifications and adaptations of the present invention are possible, and may even be desirable under certain circumstances, and they form part of the present invention. Therefore, the following description is provided to illustrate the principles of the present invention and not to limit it.

It should be understood that the expressions "for example," "such as," and variations thereof describe non-limiting embodiments of the disclosed subject matter. Reference in the description to "one exemplary embodiment," "another exemplary embodiment," "some exemplary embodiment," or variations thereof means that a particular feature, structure, or characteristic described in connection with the embodiment(s) is included in at least one embodiment of the disclosed subject matter. Thus, use of the phrase "one exemplary embodiment," "another exemplary embodiment," "some exemplary embodiment," or variations thereof does not necessarily refer to the same embodiment(s).

Unless otherwise indicated, certain features of the subject matter described herein, which for clarity are described in the context of separate embodiments, may also be provided in combination in a single embodiment. Similarly, various features of the subject matter disclosed herein, which are described in the context of a single embodiment, may also be provided separately or in any suitable subcombination.

With reference to Fig. 5 of the drawings, an example of a diamond mine data processing system is generally designated by reference numeral 12. Diamond detection is used herein as an example of implementing embodiments of the invention. However, it should be understood that alternative embodiments extend to other types of detection and image reconstruction, particularly in high-volume, high-flow environments. The concepts described herein can be applied accordingly to other tomographic imaging technologies, such as PET and/or CT, to reduce image artifacts caused by attenuation and/or scattering.

The diamond mine data processing system 12 typically includes a classification or detection system 10, which is typically an automated system capable of visualizing and detecting both individual objects within host objects and objects included in a mass of other objects.

A particular advantage of the present invention is that the system 12 is capable of producing reconstructed PET images of a rock to facilitate diamond detection.

The diamond mine data processing system 12 may be located at or near the diamond mine and may include suitable standard mining equipment, such as a crusher 14 for coarsely crushing rock mass (e.g., kimberlite) into pieces approximately 160 mm in diameter or smaller. The system 12 further comprises a suitable emitter 16 for irradiating the crushed rock with photons. The photons that irradiate the rock may originate from electron bremsstrahlung gamma-ray beams with an energy of approximately 40 MeV. Alternatively, or in addition, these photons may originate from inverse Compton scattering, a plasma wakefield device, etc. The photons operate at an energy that causes giant dipole resonance (GDR) due to a nuclear reaction between the photons and carbon in the rock.

System 12 comprises a filling apparatus 18, 19 configured to retain irradiated rock for a predetermined period of time. The irradiated rock returns to moderate levels of specific radioactivity within a few minutes, by which time the dominant residual activity is PET isotopes. Therefore, filling apparatus 18, 19 is configured to retain the irradiated rock for a retention time of 20 to 30 minutes, during which PET ^11C isotopes represent the dominant activity. Then, after the retention time has elapsed, filling apparatus 18, 19 can automatically discharge the rock.

System 12 comprises a conveyor system 20, comprising suitable conveyor belts that do not attenuate PET photons, for transporting rock into system 12 in an automated mode. Conveyor system 20 can be configured to transport rock in a rock flow within system 12 at a constant, predetermined speed, for example, 1 meter per second.

The system also comprises a detector unit 22, which is located behind the filling unit 18, 19 and next to the conveyor unit 20, in particular its belt, so as to detect PET photons emitted from it.

In one exemplary embodiment, and as shown in Figs. 5 and 6, the detector setup 22 may comprise a pair of detector arrays 22.1 and 22.2 (detector planes), which are positioned above and below the conveyor belt, respectively, so that the belt and the rock moving along it are effectively placed between them. In other embodiments, the detector setup may have a different configuration, for example, it may be substantially cylindrical with the conveyor belt positioned within the cylinder.

Array 22.1 may define an upper detector plane, which consists of a plurality of individual sensing elements or detector units, and array 22.2 may define a lower detector plane, which consists of a plurality of individual sensing elements or detector units. Arrays 22.1, 22.2 have probing axes that are substantially located transversely relative to the direction of rock movement A. In this exemplary embodiment, detector assembly 22 comprises detectors designed to detect photons. In this regard, the detectors of assembly 22 may be in the form of scintillating crystals and photomultiplier tubes (PMTs) or similar sensors with appropriate electronics.

The system 12 also comprises a suitable sorter 24, which may be a mechanical sorter 24 with electronic control, which is configured to sort potentially diamond-bearing or, in other words, diamond-containing rocks or free diamonds from potentially unproductive rocks or, in other words, rocks that do not contain diamonds.

The detection system 10 is communicatively coupled to the detector system 22 and the sorter 24 to receive classification data from the detector system 22, create reconstructed PET images, and generate appropriate control signals based on the reconstructed images to control the sorter 24 to separate diamond-bearing rocks from non-productive rocks. In this regard, it is important that the system 10 process data at a sufficient rate to be able to transmit activation signals to the sorter 24 in a timely manner.

The sorter 24 may be configured to sort diamonds or diamondiferous rocks into one or more categories according to one or more specific properties of the detected diamond/diamondiferous rock, as opposed to simply separating diamonds/diamondiferous rock from non-productive rock.

The system 10 may be connected to the detector installation 22 and/or the sorter 24 by a wired or wireless connection. In one exemplary embodiment, the system 10 is communicatively connected to the installation 22 via a communications network, which may comprise one or more different types of communications networks. In this regard, the communications network may be one network or more networks of the Internet, a local area network (LAN), a wide area network (WAN), a metropolitan area network (MAN), different types of telephone networks (e.g., a public switched telephone network (PSTN) with digital subscriber line (DSL) technology) or a mobile communication network (e.g., a global system for mobile communications (GSM), general packet radio service (GPRS), code division multiple access (CDMA), and other suitable mobile communication network technologies), or any combination thereof. It follows that, although it may not necessarily be practical, it is contemplated that in some exemplary embodiments, the system 10 need not be located on the mine site, but may be remote from it. In some embodiments, installation 22 may form part of system 10.

The system 10 is configured/programmed to receive classification data, or raw data, and to determine whether an object is potentially diamond or diamond-bearing by processing the received classification data. The classification data is typically associated with photons, detected by the installation 22, which are emitted from the object as a result of positron annihilation in the irradiated object coming from the hopper 18. In particular, with reference to Fig. 6 of the drawings, when the PET isotope in kimberlite emits a positron particle as a result of beta decay, the positron annihilates with a neighboring electron. As explained above in the "Background Art" section, the most common result of such annihilation is the formation of nearly collinear, oppositely oriented 511 eV gamma-ray photons. Each photon then passes through the surrounding material, sometimes changing energy and direction along the way. When photons reach arrays 22.1, 22.2 of the detector setup at points S1, S2, setup 22 outputs classification data in the form of collision/impact event data with the sensitive element, for example, which contains data indicating the location of the collision with array 22.1, 22.2, the photon energy and a time stamp. This makes it possible to determine or calculate the position of the diamond (see, for example, "D" in Fig. 6).

As explained above in the "Prior Art" section, a 3D model can be derived from a 4D sinogram to reconstruct the PET isotope density distribution image. This sinogram consists of the complete set of detection data acquired by detector setup 22 after sorting the response lines by angle and position. An iterative algorithm is typically used to find the most probable initial 3D source point density and, thus, to construct the 3D model.

However, based on experience using known technologies, the present inventors have found that artifacts appear in reconstructed PET images if the reconstruction process does not account for the attenuation and/or scattering of photons as they pass through the rock. Embodiments of the present invention address this problem. Various methods for compensating for attenuation and/or scattering can be used. First, mathematical algorithms can be used to model the effect using only the evolving approximate representation provided by the PET reconstruction algorithm. Second, data fusion technologies can be used to assist the PET reconstruction algorithm by obtaining additional data on the rock flow geometry obtained from other sensors. Third, gaps between or within rock layers that could lead to the occurrence or enhancement of artifacts can be filled with a special material. Fourth, artificial intelligence (AI) can be used to create images virtually free of artifacts by learning how to compensate for attenuation and scattering. Each of these methods, which may be used individually or in combination by system 10 or any other component of system 12 to reduce artifacts in PET image reconstruction, will be described below.

The first technology described in this paper is the use of algorithms that model the effects of attenuation and/or scattering using only the evolving approximate representation imparted by the PET reconstruction algorithm.

As a rough estimate, the volume imaged by system 12 shown in Fig. 5 consists of two types of material: activated rock containing the PET isotope sources and interstices between the activated rock that have no PET activity. Photons can be scattered or absorbed as they pass through the rock, but this is unlikely in the interstices, especially if these interstices are filled with air. This means that PET diamond detection, unlike PET imaging in other fields, such as medicine, has the unique characteristic that the activity of a given volume is an approximate representative variable of the probability that photons will be scattered or absorbed within that volume.

Meanwhile, it's important to understand that medical PET typically visualizes function. These corrections require specific shape inputs. There's a certain visualization of tissue proportional to its function or metabolism. However, to fully implement this correction algorithm, shape information is also required. Mineral PET (MinPET) visualizes shape because all components of the object under study contain specific isotopic species that can be activated to form a PET isotope. Therefore, all minerals in the sample will be activated to a certain degree to become PET emitters and therefore be quantitatively visualized.

Thus, correcting for scattering or attenuation in traditional PET imaging typically requires additional information about the structure and composition of the scanned objects, which can be obtained using various complementary scanning technologies, such as CT scanning (using X-rays) or magnetic resonance imaging (MRI). However, in PET imaging of diamond, the reconstructed PET image itself can be used to obtain scattering and attenuation information. When using PET in medicine, providing accompanying structural information using an alternative technology is required, as described. When improving reconstruction quality using CT, a rough 3D shape visualization is used to guide the reconstruction algorithm toward improved quality.

A PET image's initial reconstruction can reliably yield segmentation because fewer material types are involved, there is no dependence on isotope uptake (a biological function), and there is greater variability in isotope concentrations across materials. For example, a region of low isotope concentration in medical PET could represent any of several tissue types, whereas in diamond PET imaging, as discussed above, there are two primary material types in the image field: air spaces without PET activity and rock/kimberlite, which is assumed to be relatively homogeneous. If kimberlite or host rock is not substantially uniformly approximated for MinPET purposes, segmentation into additional material types can only be imagined. In the case of medical PET, shape-dependent information is acquired in a separate process, as described above, while in the case of CT, it can be obtained from a previous iteration of the CT reconstruction.

The above information can be used to implement iterative reconstruction technology. A system matrix is defined that associates each voxel in the image with each possible pair of sensing elements. Its value essentially provides the probability that positron annihilation in a given voxel will create a response line in a given pair of sensing elements. During image reconstruction, the system matrix can be iteratively updated. For example, the following steps can be performed during iterative reconstruction:

1. A sufficiently accurate approximation of the shape of the visualized object is made (for example, rock is approximately defined as a homogeneous flat layer of material). This ensures the creation of an initial image.

2. The region of interest is transformed (segmented) into a voxel-based attenuation map. Each voxel is assigned a value of 1 if it is believed to contain rock, and a value of 0 if it is believed to contain air.

3. To update the reconstructed PET image, one iterative PET reconstruction step is performed. The system matrix value for each response line is adjusted based on how many rock voxels in the attenuation map the response line passes through.

4. Each voxel in the attenuation map is updated based on the current PET image. If the PET activity exceeds a predefined threshold, the voxel is considered to contain rock and is assigned a value of 1. Otherwise, it is considered to contain air and is assigned a value of 0.

5. Then this process is repeated from step 3 above.

The voxels in the attenuation map can be the same size as the voxels in the PET image, or they can be different in size. For example, to increase computation speed, it may be necessary to use voxels from a coarser attenuation map.

In the above algorithm, as the reconstructed image converges, it becomes possible to distinguish rock contours, meaning the attenuation map becomes more accurate over time. This, in turn, improves the quality of the PET reconstruction. This technology is explained in more detail below with reference to Figs. 7-10.

The ML-EM (maximum likelihood expectation maximization) algorithm can be used for image reconstruction. ML-EM uses the timestamp of each response line to transform its position from the rock frame to the conveyor belt frame. Response lines with similar orientations (defined by the spherical coordinate system angles φ and θ) are grouped together to form a set of projections characterized by midpoint values φ and θ and two coordinates for the position along the central plane between the detectors, u and ν. To simplify notation, a single exponent j is used to label projection elements, so the set of all projections in a 4D sinogram is denoted y _i , where j is calculated over all combinations of projection variables, u, ν, φ, and θ.

The algorithm starts with an initial guess (randomly generated) regarding the distribution of PET isotopes within a given volume to be processed, denoted x _i ⁽⁰⁾ , where i is calculated over all voxels in the volume under consideration. In the simplest form of the ML-EM algorithm, image reconstruction occurs according to:

Where:

• x _i ⁽ⁿ⁾ is the value of voxel i after n iterations,

• A _ij is the “system matrix”, providing the probability that PET events in voxel i result in detection in projection element j, and

• y _j represents the detector data, grouped as described above.

Formula (2) can be understood as containing several steps. Given the current PET isotope data, x _i ⁽ⁿ⁾ , the data are forward projected to see what would be found if this were the true PET isotope distribution for a given projection element j. The forward projection uses the system matrix, A:

Then the directly projected data z _j are compared with the actual data y _j using the relation:

The correction for a given voxel is then calculated by adding the corrections for each projection element j:

In formula (3) above, the term in parentheses represents the normalization factor. The sum term on the right is the back-projection of the ratio of the true detected data to the forward-projected data. This summed correction is then applied to the current guess x _i ⁽ⁿ⁾ to obtain the next guess x _i ⁽ⁿ⁺¹⁾ . This step is known as "back-projection."

The success of ML-EM technology lies in its ability to accurately perform forward and back projection. In other words, given a specific isotope and kimberlite distribution, it's possible to obtain a good approximation of what will be captured by the detector array's sensitive elements. These corrections can then be accurately propagated back to the image for the next iteration.

In the simplest reconstruction procedure, the system matrix _Aij is zero everywhere except where a given voxel i is directly in the path of projection element j. This does not converge into an exact image, since what is actually detected is influenced by factors such as:

• attenuation, which reduces the number of response lines in a projection element in proportion to the amount of absorbing material along that projection element;

• random coincidences that lead to background response lines throughout the system; and

• scattering, which results in events occurring in one projection element being detected in another.

To accurately generate a reconstructed PET image, equation (2) must be updated to account for these effects. There are several ways to do this. One is to "correct" the detected data to calculate what would have been detected in the absence of these effects. For example, equation (2) can be expanded to:

Where:

• R _j is the expected number of random matches detected by projection element j;

• S _j is the expected number of random scattered response lines detected by projection element j;

• T _j is a coefficient that records how much the number of response lines in projection element j has decreased due to attenuation; and

• N _j - normalization correction.

Another way to conceptualize the scattering and attenuation adjustments is to introduce these effects into the system matrix themselves:

Where:

• S _ij denotes the probability that a source isotope at voxel i will result in a detection within projection element j; and

• T _ij represents the attenuation factor that reduces the probability of detecting a PET event from voxel i within projection element j.

As explained above, a response line is created only if both oppositely oriented photons reach their respective detectors on both sides. Therefore, every point in space between two sensitive elements must be intersected by one or the other photon. Therefore, the attenuation along a given projection does not depend, to a first approximation, on the position of the source along the projection. Thus, T is effectively only a function of projection element j, not of which voxel is located along it:

The system matrix can, in principle, be extremely large. For example, a 20 cm x 20 cm x 20 cm volume with a resolution of 2 mm contains 1,000,000 voxels. A 4D sinogram may have 4,000 different angle sets, with u and v values reaching up to 100, for a total of 40,000,000 projection elements. Thus, in this case, the system matrix has a domain of 4 x ¹⁰ values. Typically, such a volume cannot be stored and is too computationally expensive to process if each potential value is generated in real time. Therefore, iterative technologies rely on the fact that most A _ij values are zero, which reduces the number of calculations. Scattering can occur at any angle, so, in principle, all S _ij values are nonzero. In order for the calculation to remain tractable, it needs to be simplified, for example by considering only the average scattering rate, which depends on the angle, in which case S _ij can be treated similarly to R _j in equation (4), or by considering only the scattering to nearby projection elements, in which case S _ij serves to "expand" A _ij by increasing the set of voxels that participate in the operation of a given sensor.

It is important to consider the distribution of absorbing material in the image, which determines _Tj and may contribute to _Sij (depending on how scattering is modeled). The term "attenuation" is used here in the broad sense of the probability with which both photons from an annihilation event should reach the detectors on each side without significant changes, for example, in direction. This is significantly broader than the traditional use of the term "linear attenuation," which refers to whether a particle is absorbed or not. To define attenuation, the concept of an "attenuation map" is introduced. This map represents a 3D image of the object's volume, where each voxel contains information about how the material within that voxel is attenuated.

In PET imaging of diamonds, an advantage is that the volume under consideration can be divided, to a preliminary approximation, into kimberlite or air (diamonds are small enough that this is not taken into account for this calculation). Only one of these, kimberlite, is a potential source of PET activity. Thus, the attenuation T _j can be calculated using the information contained in the PET image. An example of the general scheme is shown in flowchart 30 in Fig. 7. The blocks in solid lines indicate steps that are unique to the method according to embodiments of the invention, while the blocks in dashed lines are part of the standard ML-EM procedure.

As explained above, the algorithm starts with an initial guess (randomly generated) regarding the distribution of PET isotopes within the given volume to be processed (see step 32).

Next, in step 34, image segmentation is performed as described above to determine the attenuation map. Figure 8 shows examples of step 34 of image segmentation. In Figure 8, column A represents a reconstructed PET image, which is created iteratively by converging from the original noise into an image of three rocks. It should be understood that this is not an actual PET image (since an actual image is 3D), but a simulated 2D image for illustrative purposes. The remaining four columns show various approaches to segmenting this image to obtain the attenuation map.

Column B represents perhaps the simplest approach, where a threshold is calculated and each voxel is classified as kimberlite (colored black) if it exceeds the threshold, and otherwise classified as air (colored white). The resulting attenuation map is relatively noisy and does not accurately represent objects in the image, even after four iterations, as can be seen from column B in Fig. 8.

Column C shows the downsampled attenuation map. In this case, voxels are downsampled by a factor of 4, meaning each 4x4 region of the original PET image results in one voxel in the attenuation map. A similar threshold filter is then applied to the downsampled map. This provides two advantages. First, downsampling helps remove some noise compared to the images in the second column. Second, the lower-resolution attenuation map allows for faster attenuation correction calculations in the system matrix.

The method shown in column D uses more complex filters. It first applies two Gaussian blurs to the image, then applies thresholding. Finally, the method shown in column E uses the same Gaussian blurring procedure, but then downsamples as in column C before applying thresholding. Other prior art image segmentation techniques can be used. Clearly, more complex techniques should perform better, but this comes at the expense of computation time.

For each value j, the attenuation map is used to calculate the value T _j , which, as explained above, is an attenuation factor that captures how much the number of response lines in projection element j has been reduced due to attenuation (step 36).

In order to update the system matrix (step 38 in Fig. 7), an attenuation correction can be applied to it. This can be applied either according to the approach of equation (4) or equation (5).

Figure 9 shows a 2D analog of a simple algorithm for performing the attenuation calculation for a given projection. The pixels shown are a close-up of the lower right attenuation map shown in Figure 8. A line is drawn along each projection. The 3D attenuation map is then split into x-y layers along the detector connection direction z. For each layer, the voxel closest to the line is found by finding the voxel center with the closest horizontal distance to the x,y position of the line in that layer. The total number of voxels segmented into the rock is then counted. For projection element 1 in Figure 9, this total is 14, and for projection element 2, it is 11. A correction is then applied to this counted number for slanted lines. If the angle between a given projection element and the vertical is q, then the counted number of voxels must be multiplied by 1/cos(q). The adjusted voxel count is approximately equal to the path length through the kimberlite for the projection.

where p is the path length in cm, N is the number of voxels crossed, and w is the voxel width.

Once the path length through kimberlite is known, the photon attenuation can be calculated. The standard formula for linear attenuation is:

where I ₀ is the uncorrected photon intensity, I is the attenuation-corrected photon intensity, and am is the linear attenuation coefficient for a given material.

Both absorption and scattering can prevent the detection of an accurate response line. This is known as "attenuation" in the context of narrow beam measurements. The attenuation of photons within the rock can be measured experimentally. Another possible method is to place a point source containing the PET isotope between two small sensitive elements and measure the response line detection rate. Flat pieces of kimberlite of known width are then placed successively in the response line path, allowing the response line event rate to be plotted against kimberlite width. These data can be fitted to a function of the form (6) to derive the effective attenuation coefficient m _eff . For each projection element j, the path length p _j is then found according to equation (6), and the attenuation T _j is then

If the exponential decay function is inconsistent with the data when plotting the above graph, other functional forms can be used to relate the path length to the detection frequency of the response line. This can also be done computer-based, for example, using Geant4 simulations of a gradual increase in rock diameter. In addition to quantifying attenuation, this experiment is a useful way to compare Geant4 simulation results with respect to the precise manipulation of photon transport through kimberlite.

It's important to understand that more complex path length calculation methods are possible. For example, instead of calculating a single voxel, all voxels intersected by the projection can be considered, and the projection path length through each voxel is calculated to obtain a more accurate result. This calculation can be performed for each of a large number of projections per iteration, per volume, in real time. Therefore, the chosen algorithm must consider computation time as well as accuracy; therefore, a faster algorithm may be preferable to a more accurate one, depending on the technical requirements of a specific application.

An alternative method for determining path length is shown in Fig. 10. Image A in Fig. 10 is the bottom image from column A in Fig. 8. A boundary detection algorithm is used to detect kimberlite boundaries, as shown in Fig. B. In this case, the Canny algorithm was used to detect boundaries. The intersection points between the projections and the boundaries (shown as dots on the rock boundaries in Fig. C) are found and used to calculate the path length through the kimberlite. This is done in three dimensions.

Combinations of different technologies can also be used. For example, edge detection algorithms may fail to detect boundaries that completely enclose volumes in low-quality images, such as those available with a small number of projections. For example, volume segmentation technology can be used with a small number of iterations, followed by edge detection-based technology with a higher number of iterations.

This method can be further generalized to more complex attenuation map segmentation based on the presence of the most common materials on the conveyor belt. For example, instead of a binary segmentation into kimberlite and air, a four-way segmentation can be used to create an attenuation map, distinguishing between air, two types of kimberlite, and host rock, each with a different attenuation coefficient. This is only feasible if the PET concentrations of the various materials differ significantly enough to allow material differentiation.

After updating the system matrix, the data is forward projected (step 40 in Fig. 7) as described above and then back projected (step 42) to calculate an updated source volume serving as the next guess, and this process continues iteratively in the same manner.

Scattering can be modeled by including elements in the system matrix A _ij that are nonzero for nearby source voxels, but not along a direct path along a given projection. This accounts for the fact that a change in photon direction due to scattering can result in a reconstructed response line that does not pass through the source point. The scattering probability is proportional to the path length through the kimberlite. Therefore, scattering correction can be performed in a manner similar to attenuation correction, namely, first calculating the path length and then introducing nonzero elements of the system matrix separately for each projection.

It should be understood that there are many variations and extensions of the ML-EM scheme described above, such as OS-EM (ordered subset expectation maximization) and list-mode reconstruction, which processes individual response lines instead of grouping them into projections, favoring increased accuracy at the expense of processing a much larger data set. The key aspects of using PET data to create an attenuation map during the iteration process are not limited to the basic ML-EM scheme but can be applied in many iterative reconstruction algorithms in a similar manner.

As mentioned above, data fusion technologies can be used to assist the PET reconstruction algorithm by obtaining additional flow geometry data from other sensors.

Other data may be available to assist in creating the attenuation map. This additional data may either replace the algorithm defined above or complement it. Two examples of additional information include video camera data and laser cloud reflectance scanning. If the incoming material is scanned before or after it enters the detector system, at least some information about the shape of the kimberlite particles is provided. This information may be incomplete or imperfect, for example, if it is impossible to determine the presence of cavities within the kimberlite that are completely hidden from external view. However, a kimberlite surface map may contain information that can be used to create or assist in creating an attenuation map.

If an external video signal is used to create a surface map, it may involve merging multiple external video signals from different angles. Re-creating object shapes from video information obtained at different angles is a well-studied area, and various known methods can be used.

One option for augmenting the attenuation map obtained from a PET image with surface information would be to use the surface information to segment all regions outside the surface, such as air, and then use the PET image to perform segmentation within the volume enclosed by that surface.

Another source of kimberlite information that can be used is the mass scale. Kimberlite density is relatively well understood. Therefore, a position-dependent scale can provide real-time information on the total volume of kimberlite present in a given area. This information, either alone or in combination with the surface information described above, can be used to constrain potential attenuation maps.

Another technique for reducing image artifacts is to fill the spaces between the analyzed objects. There are many ways to fill the voids between rocks. One method involves pouring a liquid into the kimberlite layer. The kimberlite layer is allowed to settle, then passed through detectors, and the liquid is withdrawn on the other side and reused. Water can be used, but its attenuation is much less than that of kimberlite, so more accurate results can be obtained using denser liquids or suspended particles. An iron-siliceous suspension, whose density exceeds that of kimberlite, can be used. Kimberlite powder can also be used. In this case, the likelihood of filling all the cavities is lower, but this powder is present in large quantities in diamond mines, does not negatively impact rock stability, and does not require extensive water disposal circuits. Fine particles (kimberlite particles are less than 2 mm in diameter) are commonly generated during rock crushing and transport and loading operations. These are typically removed using standard diamond detection methods. However, they can be preserved and used for the purposes mentioned above. A vibrating conveyor belt can act as a source of fine particles with liquid properties, which can then quickly settle into the cavities between larger rocks.

Although filling cavities with additional material does not completely reduce scattering, as the new material also scatters photons, it does reduce scattering variations at different positions within the rock. Therefore, this technology could potentially also help reduce artifacts in the image.

Finally, artificial intelligence (AI) can be used to create images virtually free of artifacts by learning how to compensate for attenuation and/or scattering. For example, a deep learning AI system based on a convolutional neural network (CNN) can learn how to remove the effects of attenuation and scattering from a reconstructed PET image during post-imaging processing. The CNN can be trained using Geant4-simulated data on various rock phantoms to create realistic images (with attenuation and scattering), which, along with ideal images (assuming response lines with equal activity without scattering/attenuation), can represent idealized geometric shapes, forming a "true" PET image. When PET is used in medicine, these operations cannot be implemented. In this case, there is a large increase in the number of cells of the possible tissue types present, with no obvious relationship between the PET source intensity and photon attenuation. Furthermore, attenuation in biological tissue is less than in rock. Thus, scattering and attenuation artifacts in medical images are in most cases subtle, complex, highly tissue-dependent, and indistinguishable from image features.

On the other hand, in PET diamond detection, with a simplified set of materials with a clear relationship between source intensity and attenuation, as well as with a sharp attenuation contrast between rock and air, defects caused by attenuation and scattering are large and distinct from image features. Therefore, a generative neural network can learn the characteristics of such artifacts and create an image that corresponds to what an image without scattering and attenuation should look like. Again, the ability to create Geant4 data with "true" information is an advantage, as without it, it would be difficult to obtain images without attenuation and scattering from experiments with real data for training the neural network.

In another embodiment, instead of applying a neural network to an already-created image, it can be used in the iterative reconstruction process. Specifically, the neural network can be trained to generate an attenuation map from the current source concentration map, instead of the various thresholding and segmentation techniques described above.

Referring now to Fig. 11 of the drawings, a schematic representation of a machine, in the example of a computer system 100, is shown, in which a set of instructions may be executed that cause the machine to perform any one or more of the techniques discussed herein. In other exemplary embodiments, the machine operates as a stand-alone device or may be connected (e.g., via a network) to other machines. In a networked exemplary embodiment, the machine may operate as a server or a client machine in a server-client network environment, or as a peer machine in a peer-to-peer (or distributed) network environment. The machine may be a personal computer (PC), a tablet PC, a set-top box (STB), a personal digital assistant (PDA), a cellular telephone, a web device, a network router, switch or gateway, or any other device capable of executing a set of instructions (sequential or otherwise) that specify actions to be performed by the machine. Furthermore, although only one machine is shown for convenience, the term "machine" should also be understood to include any collection of machines, including virtual machines, that individually or collectively execute a set (or multiple sets) of instructions to implement any or more of the techniques discussed herein.

In any case, the exemplary computer system 100 includes a processor 102 (e.g., a central processing unit (CPU), a graphics processing unit (GPU), or both), a main memory 104, and a static memory 106 that communicate with each other via a bus 108. The computer system 100 may further include a video display unit 110 (e.g., a liquid crystal display (LCD) or a cathode ray tube (CRT)). The computer system 100 also includes an alphanumeric input device 112 (e.g., a keyboard), a user interface (UI) navigation device 114 (e.g., a mouse or a touchpad), a disk drive unit 116, a signal supply device 118 (e.g., a speaker), and a network interface device 120.

The disk drive 16 includes a non-volatile computer-readable medium 122 storing one or more sets of instructions and data structures (e.g., software 124) implementing any one or more of the techniques or functions described in this document or used by them. The software 124 may also reside, in whole or at least in part, in the main memory 104 and/or in the processor 102 during its execution by the computer system 100, the main memory 104 and the processor 102 also constituting the computer-readable medium.

The software 124 may further be transmitted or received over the network 126 using the network interface device 120 using any of a number of well-known transmission protocols (e.g., HTTP).

Although the computer-readable medium 122 is shown in the exemplary embodiment as a single medium, the term "machine-readable medium" may refer to a single medium or multiple media (e.g., a centralized or distributed data storage and/or associated buffer memory and servers) that store one or more instruction sets. The term "machine-readable medium" may also be considered to include any medium that is capable of storing, encoding, or carrying a set of instructions for execution by a machine, which instruct the machine to implement one or more methodologies of the present invention, or that is capable of storing, encoding, or carrying data structures used by or associated with such a set of instructions. Accordingly, the term "machine-readable medium" may include, but is not limited to, solid-state memory, optical and magnetic media, and carrier wave signals.

The present inventors believe that embodiments of the present invention can effectively reduce artifacts in PET image reconstruction. In particular, embodiments of the present invention can enable the system to correct for scattering and/or attenuation during the iterative image reconstruction process when using PET for mineral detection.

As mentioned above, in the "Background Art" section, we used PET enhancement in a mining environment (MinPET) to explain the present invention for attenuation and scatter correction during image reconstruction. However, those skilled in the art will readily appreciate that the concepts of the present invention are equally applicable to PET and CT scanning in medicine.

Claims (41)

1. A method for reconstructing positron emission tomography (PET) images, wherein the method is carried out in whole or in part by a computer or an automated system connected in communication with a detector system that contains a plurality of sensitive elements, wherein the sensitive elements are configured to detect photons associated with an object, and the method comprises the following steps: creating or obtaining an initial reconstruction of a PET image associated with an object, wherein the PET image is at least partially based on lines of response (LoRs) detected by a detector system, wherein the PET image includes a plurality of voxels and provides a representation of the density distribution of the PET isotope in a volume or region analyzed by the detector system; defining or obtaining a system matrix that relates each voxel in the PET image to the sensitive elements; creating or obtaining an attenuation map, wherein the attenuation map is based on a PET image and is defined as a plurality of voxels, wherein each voxel of the attenuation map contains information about the voxel concerning the degree or probability of attenuation and/or scattering associated with that voxel; performing one iteration of the iterative PET image reconstruction algorithm to update the PET image, wherein performing the iteration includes updating the system matrix based on the attenuation map voxel information; updating the attenuation map based on the updated PET image; and repeating one iteration of the reconstruction algorithm and updating the attenuation map until the image converges to determine a reconstructed PET image. 2. The method of claim 1, wherein the method comprises using a machine learning-trained imaging device to create a reconstructed PET image, wherein the imaging device is trained on at least computer-simulated images and/or object data obtained from a computerized simulator to compensate for or take into account attenuation and/or scattering of photons associated with the object, thereby reducing artifacts in the reconstructed image. 3. The method according to claim 1 or 2, wherein the method comprises a first step of filling the cavities between the objects with a predetermined material or materials before performing analysis of the object(s) by the detector system. 4. The method of any one of the preceding claims, wherein the method includes using additional data from the sensor to create, obtain and/or update the attenuation map. 5. The method of claim 4, wherein the additional data from the sensor is based on data from a video camera and/or laser reflection scanning in the cloud. 6. The method of any one of the preceding claims, wherein the method is used to reduce artifacts in PET image reconstruction by compensating for or taking into account attenuation and/or scattering of photons emitted from the object. 7. A method according to any one of the preceding claims, wherein the sensitive elements are configured to detect photons emitted from the object as a result of positron annihilation caused by irradiating the object with photons of a predetermined energy, wherein the PET image provides a representation of the density distribution of the PET isotope in the volume or region analyzed by the detector apparatus. 8. The method of claim 7, wherein the system matrix associates each voxel in the PET image with each possible pair of sensitive elements. 9. A system for reducing artifacts in positron emission tomography (PET) image reconstruction, wherein the system comprises a computer or automated system communicatively coupled to a detector system that comprises a plurality of sensitive elements, wherein the sensitive elements are configured to detect photons associated with an object, wherein the computer or automated system is configured to: creating or obtaining an initial reconstruction of a PET image associated with an object, wherein the PET image is at least partially based on lines of response (LoRs) detected by a detector system, wherein the PET image includes a plurality of voxels associated with a volume or region analyzed by the detector system; determining or obtaining a system matrix that relates each voxel in the PET image to the sensitive elements; creating or obtaining an attenuation map, wherein the attenuation map is based on the PET image and is defined as a plurality of voxels, wherein each voxel of the attenuation map contains information about the voxel concerning the degree or probability of attenuation and/or scattering associated with that voxel; performing one iteration of the iterative PET reconstruction algorithm to update the PET image, wherein performing the iteration includes updating the system matrix based on the attenuation map voxel information; updating the attenuation map based on the updated PET image; and repeating one iteration of the reconstruction algorithm and updating the attenuation map until the image converges to determine a reconstructed PET image. 10. The system of claim 9, wherein the computer or automated system is configured to use a machine learning-trained imaging device to create a reconstructed PET image, wherein the imaging device is trained on at least computer-simulated images and/or object data obtained from a computerized simulator to compensate for or take into account attenuation and/or scattering of photons associated with the object, thereby reducing artifacts in the reconstructed image. 11. The system of claim 9 or 10, wherein the computer or automated system is configured to use additional data from the sensor to create, receive and/or update the attenuation map. 12. The system of claim 11, wherein the additional data from the sensor is based on data from a video camera and/or laser reflection scanning in the cloud. 13. The system of any one of claims 9-12, wherein the system is configured to reduce artifacts in the reconstruction of a positron emission tomography (PET) image by compensating for or taking into account the attenuation and/or scattering of photons emitted from the object. 14. The system of any one of claims 9 to 13, wherein the sensitive elements are configured to detect photons emitted from the object as a result of positron annihilation caused by irradiating the object with photons of a predetermined energy, wherein the PET image provides an idea of the density distribution of the PET isotope in the volume or region analyzed by the detector system. 15. The system of claim 14, wherein the system matrix associates each voxel in the PET image with each possible pair of sensitive elements. 16. The system of any one of claims 9-12, wherein the system is configured to reduce artifacts in PET image reconstruction by compensating for or taking into account scattering of photons from the source as it interacts with the object. 17. The system according to any one of paragraphs 9-16, in which the system comprises a detector installation. 18. The system of any one of claims 9 to 17, wherein the system is part of a diamond mine data processing system. 19. A method for reconstructing positron emission tomography (PET) images, wherein the method is performed by a computer or an automated system communicatively coupled to a detector system that contains a plurality of sensing elements, wherein the sensing elements are configured to detect photons associated with an object, wherein the method includes using a machine learning-trained imaging device to create a reconstructed PET image associated with the object, wherein the imaging device is trained on at least computer-simulated images and/or object data obtained from a computerized simulator to compensate for or take into account attenuation and/or scattering of photons associated with the object. 20. The method of claim 19, wherein the method is used to reduce artifacts in PET image reconstruction by compensating for or taking into account attenuation and/or scattering of photons emitted from the object. 21. The method according to claim 19 or 20, wherein the sensitive elements are configured to detect photons emitted from the object as a result of positron annihilation caused by irradiating the object with photons of a predetermined energy, wherein the PET image provides an idea of the density distribution of the PET isotope in the volume or region analyzed by the detector system. 22. The method of any one of claims 19-21, wherein the method is used to reduce artifacts in PET image reconstruction by compensating for or taking into account the scattering of photons from the source as it interacts with the object. 23. The method according to claim 19 or 22, wherein the sensitive elements are configured to detect photons from a source passing through the object, wherein the PET image provides an indication of the material density and/or electron density in the object analyzed by the detector system. 24. A system for reconstructing positron emission tomography (PET) images, wherein the system comprises a computer or an automated system communicatively coupled to a detector assembly that comprises a plurality of sensing elements, wherein the sensing elements are configured to detect photons associated with an object, wherein the computer or automated system is configured to detect photons associated with an object, wherein the system is configured to use a machine learning-trained imaging device to create a reconstructed PET image associated with an object, wherein the imaging device is trained on at least computer-simulated images and/or object data obtained from a computerized simulator to compensate for or take into account attenuation and/or scattering of photons associated with the object. 25. The system of claim 24, wherein the system is configured to reduce artifacts in the reconstruction of a positron emission tomography (PET) image by compensating for or taking into account the attenuation and/or scattering of photons emitted from the object. 26. The system of claim 24 or 25, wherein the sensitive elements are configured to detect photons emitted from the object as a result of positron annihilation caused by irradiating the object with photons of a predetermined energy, wherein the PET image provides an indication of the density distribution of the PET isotope in the volume or region analyzed by the detector system. 27. The system of claim 26, wherein the system matrix associates each voxel in the PET image with each possible pair of sensitive elements. 28. The system of claim 24, wherein the system is configured to reduce artifacts in PET image reconstruction by compensating for or taking into account scattering of photons from the source as it interacts with the object. 29. The system according to any one of paragraphs 24-28, wherein the system comprises a detector installation. 30. The system according to any one of paragraphs 24-28, 29, wherein the system is part of a diamond mine data processing system. 31. A method for reconstructing positron emission tomography (PET) images, wherein the reconstructed PET image is generated by a computer or automated system communicatively coupled to a detector system that contains a plurality of sensing elements, wherein the sensing elements are configured to detect photons associated with objects, the method comprising filling cavities between objects or in an object(s) with a predetermined material or materials prior to analysis of the objects by the detector system.