Information processing device, information processing method, pet device, program, and recording medium

Abstract

The present invention realizes a technique for performing crystal identification in a plurality of radiation detectors without using pLUT during operation. A processor (11) executes crystal identification processing for identifying, by using a single model (μ) generated by machine learning, a crystal in which photoelectric absorption has occurred in a radiation detector in which an event has occurred from among a plurality of radiation detectors each including a crystal array and a light-receiving element array. The model (μ) is a model in which an output signal of the light-receiving element array included in the radiation detector is used as an input, and a likelihood corresponding to each crystal constituting the crystal array included in the radiation detector and being a likelihood that the crystal is a crystal in which photoelectric absorption has occurred is used as an output.

Description

Information processing device, information processing method, PET device, program, and recording medium

[0001] This disclosure relates to an information processing device and an information processing method for performing a crystal identification process to identify crystals that have undergone photoelectric absorption in a radiation detector. It also relates to a PET apparatus equipped with such an information processing device, a program for causing a computer to execute the information processing method, and a recording medium.

[0002] A radiation detector used in PET (Positron Emission Tomography) systems and the like comprises a crystal array consisting of multiple crystals arranged in a matrix, and a photodetector array consisting of multiple photodetectors arranged in a matrix. When radiation is incident on the radiation detector, photoelectric absorption occurs in the incident crystal, causing fluorescence to be emitted. The photodetector array outputs a signal representing the fluorescence intensity detected by each photodetector at this time.

[0003] In conventional radiation detectors, a centroid calculation is performed to determine the centroid position of the fluorescence intensity based on the spatial extent of the fluorescence intensity detected by each photodetector, in order to identify which crystal in the crystal array is causing photoelectric absorption. An example of a centroid calculation is shown in Figure 8. Figure 8 illustrates the centroid calculation in a radiation detector equipped with a crystal array consisting of 6x6 crystals and a photodetector array consisting of 2x2 photodetectors.

[0004] Identifying crystals that exhibit photoelectric absorption is performed by comparing the centroid position obtained through centroid calculation with a lookup table called a "position look-up table (pLUT)". The pLUT is a table that associates the address of a crystal that exhibits photoelectric absorption with the centroid position obtained through centroid calculation. The pLUT can be visualized as a two-dimensional map. An example of a two-dimensional map representing the pLUT is shown in Figure 9. In Figure 9, the pLUT of a radiation detector equipped with a crystal array consisting of 6x6 crystals and a photodetector array consisting of 2x2 photodetectors is shown as an example.

[0005] Each region A shown in the 2D map ｉ，ｊThis is the region corresponding to the crystal with address (i, j) (the crystal arranged in row i and column j). For example, the centroid position (X, Y) obtained by centroid calculation corresponds to region A. １，１ If included in region A, the crystal with address (1,1) is identified as the crystal where photoelectric absorption occurred. Alternatively, if the centroid position (X,Y) obtained by centroid calculation is in region A １，２ If included in the region, the crystal with address (1,2) is identified as the crystal where photoelectric absorption occurred. Generally, the centroid position (X,Y) obtained by centroid calculation is in region A ｉ，ｊ If included in the set, the crystal with address (i, j) is identified as the crystal that has undergone photoelectric absorption.

[0006] Furthermore, Compton scattering can occur when radiation is incident on a radiation detector. When radiation is emitted outside the crystal array due to Compton scattering, the variation in the centroid position increases, but the radiation can be associated with any crystal (i.e., the address of a crystal). On the other hand, when Compton-scattered radiation is photoelectrically absorbed within the crystal array (also called "intra-detector scattering"), the centroid position obtained by centroid calculation becomes a centroid position that cannot be associated with any crystal address in the pLUT, and the spatial resolution deteriorates. In this case, since fluorescence is detected in a dispersed manner, the temporal resolution also deteriorates. Therefore, (1) based on whether the centroid position obtained by centroid calculation is associated with any crystal (address) in the pLUT, it is possible to identify whether the event that occurred is photoelectric absorption or intra-detector scattering, and if the event that occurred is photoelectric absorption, (2) based on which crystal (address) the photoelectric absorption occurred is identified based on which crystal the centroid position obtained by centroid calculation is associated with in the pLUT. In event identification processing using pLUT, the photoelectric absorption region (A ｉ，ｊ Intracrystalline scattering that overlaps with the photoelectric absorption region (A) is difficult to distinguish. Therefore, in order to reduce or eliminate intracrystalline scattering, the photoelectric absorption region (A) is used. ｉ，ｊ It is necessary to make this as small as possible (minimize it), but this minimization leads to a decrease in sensitivity to radiation.

[0007] Hereinafter, the process of identifying whether an event occurring in a radiation detector is photoelectric absorption or scattering within the detector will be referred to as "event identification," and the process performed by a processor in a computer or similar device to perform event identification will be referred to as "event identification processing." Similarly, the process of identifying a crystal in which photoelectric absorption has occurred in a radiation detector will be referred to as "crystal identification," and the process performed by a processor in a computer or similar device to perform crystal identification will be referred to as "crystal identification processing."

[0008] Examples of literature concerning crystal identification in radiation detectors that constitute PET devices include the following:

[0009] Yoshida E, et al, 2024 Timing estimation of the exponentiated energy-weighted average for crosshair light sharing TOF-DOI PET detector Nucl, Instrum. Methods Phys. Res. Sect. A: Accel., Spectrometers, Detect. Assoc. Equip. 1059 168949

[0010] Multiple radiation detectors are used in PET scanners and similar devices. These detectors can exhibit variations in their 2D maps due to factors such as crystal surface conditions, fluorescence intensity variations, crystal array fabrication accuracy, and individual differences in the detection elements. Furthermore, PET scanners targeting small animals or human heads require even finer crystals, necessitating highly accurate crystal identification. Therefore, manufacturing a PET scanner requires the creation of a pLUT (ply-luminescent spectroscopy unit) for each radiation detector. Additionally, radiation detector characteristics change over time. Consequently, the pLUTs must be periodically updated to operate the PET scanner. The process of creating and updating the same number of pLUTs as the radiation detectors contributes to increased manufacturing and operating costs for the PET scanner.

[0011] One aspect of the present invention has been made in view of the above-mentioned problems, and its objective is to realize a technology for performing crystal identification in multiple radiation detectors without using a pLUT during operation.

[0012] An information processing device according to one aspect of the present invention includes a processor, which performs a crystal identification process to identify, using a single model generated by machine learning, the crystal in which photoelectric absorption occurred in a radiation detector where an event occurred, from among a plurality of radiation detectors, each including a crystal array and a photodetector array, wherein the model takes the output signal of the photodetector array included in the radiation detector as input and outputs the likelihood corresponding to each crystal constituting the crystal array included in the radiation detector, which is the likelihood that the crystal is a crystal in which photoelectric absorption occurred.

[0013] An information processing apparatus according to one aspect of the present invention comprises a processor, the processor performing a centroid calculation process on the output signal of a photodetector array included in a plurality of radiation detectors, each including a crystal array and a photodetector array, where an event has occurred; an event identification process that identifies whether the event is photoelectric absorption or in-crystal scattering by comparing the centroid position calculated in the centroid calculation process with the pLUT; and if the event is identified as photoelectric absorption in the event identification process, the processor compares the centroid position calculated in the centroid calculation process with the pLUT to determine whether the event is photoelectric absorption. The following steps are performed: a crystal identification process that identifies the crystal that has undergone photoelectric absorption in the radiation detector where the event occurred; a training data generation process that generates training data by adding the address of the crystal identified in the crystal identification process as a correct label to the output signal of the photodetector array included in the radiation detector where the event occurred; and a model learning process that uses the training data generated in the training data generation process to learn a model that takes the output signal of the photodetector array included in the radiation detector as input and outputs the likelihood corresponding to each crystal constituting the crystal array included in the radiation detector, where the crystal is the crystal that has undergone photoelectric absorption.

[0014] An information processing method according to one aspect of the present invention includes a crystal identification process in which a processor identifies a crystal that has undergone photoelectric absorption in a radiation detector where an event has occurred, using a single model generated by machine learning, from among a plurality of radiation detectors, each of which includes a crystal array and a photodetector array, wherein the model takes the output signal of the photodetector array included in the radiation detector as input and outputs the likelihood corresponding to each crystal constituting the crystal array included in the radiation detector, which is the likelihood that the crystal has undergone photoelectric absorption.

[0015] An information processing method according to one aspect of the present invention includes: a centroid calculation process in which a processor performs a centroid calculation on the output signal of a photodetector array included in a plurality of radiation detectors, each of which includes a crystal array and a photodetector array, where an event has occurred; an event identification process in which the processor identifies whether the event is photoelectric absorption or in-crystal scattering by comparing the centroid position calculated in the centroid calculation process with a pLUT; and if the event is identified as photoelectric absorption in the event identification process, the processor identifies whether the event has occurred by comparing the centroid position calculated in the centroid calculation process with the pLUT. The process includes: a crystal identification process that identifies a crystal in which photoelectric absorption has occurred in a radiation detector; a training data generation process in which the processor generates training data by adding the address of the crystal identified in the crystal identification process as a correct label to the output signal of the photodetector array included in the radiation detector in which the event occurred; and a model learning process in which the processor learns a model using the training data generated in the training data generation process, taking the output signal of the photodetector array included in the radiation detector as input and outputting the likelihood corresponding to each crystal constituting the crystal array included in the radiation detector, which is the likelihood that the crystal is a crystal in which photoelectric absorption has occurred.

[0016] According to one aspect of the present invention, crystal identification in multiple radiation detectors can be performed without using a pLUT during operation.

[0017] Figure 1 is a block diagram showing the configuration of an information processing device according to one embodiment of the present invention. Figure 1 is a flowchart showing the flow of the information processing method performed by the information processing device shown in Figure 1 during the inference phase. Figure 1 is a schematic diagram showing a specific example of a model used by the information processing device shown in Figure 1. Figure 1 is a flowchart showing the flow of the information processing method performed by the information processing device shown in Figure 1 during the learning phase. For each of the 20 events in which the crystal identification result using the model shown in Figure 3 was correct, and the 10 events in which the crystal identification result using the model shown in Figure 3 was incorrect, the images represent the input of model μ (a set of fluorescence intensities detected by each photodetector) and the output of model μ (a set of likelihoods that each crystal is a crystal in which photoelectric absorption occurred). (a) is a timing histogram generated by referring to the address of the crystal obtained by the information processing method according to the comparative example, and (b) is a timing histogram generated by referring to the address of the crystal obtained by the information processing method shown in Figure 2. (a) is a reconstructed image generated by referring to the address of the crystal obtained by the information processing method according to the comparative example, and (b) is a reconstructed image generated by referring to the address of the crystal obtained by the information processing method shown in Figure 2. This is a schematic diagram illustrating a specific example of centroid calculation in conventional crystal identification. This image shows a specific example of a pLUT used in conventional crystal identification.

[0018] (Configuration of Information Processing Device) The configuration of the information processing device 1 according to one embodiment of the present invention will be described with reference to Figure 1. Figure 1 is a block diagram showing the configuration of the information processing device 1.

[0019] The information processing device 1 is implemented using a general-purpose computer and, as shown in Figure 1, comprises a processor 11, a primary memory 12, a secondary memory 13, an input / output interface 14, and a bus 15. The processor 11, primary memory 12, secondary memory 13, and input / output interface 14 are interconnected via the bus 15. The secondary memory 13 stores (non-volatile) information processing programs P1 and P2.

[0020] During the learning phase, the processor 11 loads the information processing program P2 stored in the secondary memory 13 onto the primary memory 12. Then, the processor 11 executes each process included in the information processing method S2, which will be described later, according to the information processing program P2 loaded onto the primary memory 12. By executing each process included in the information processing method S2, the processor 11 generates the model μ, which will be described later. The generated model μ is stored in the secondary memory 13.

[0021] In the inference phase, the processor 11 loads the information processing program P1 and model μ stored in the secondary memory 13 onto the primary memory 12. Then, the processor 11 executes each process included in the information processing method S1, which will be described later, according to the information processing program P1 loaded onto the primary memory 12. When executing each process included in the information processing method S1, the processor 11 utilizes the model μ loaded onto the primary memory 12.

[0022] An example of a device that can be used as the processor 11 is a CPU (Central Processing Unit). An example of a device that can be used as the primary memory 12 is semiconductor RAM (Random Access Memory). An example of a device that can be used as the secondary memory 13 is an HDD (Hard Disk Drive).

[0023] Input and / or output devices are connected to the input / output interface 14. An example of an input device connected to the input / output interface 14 is a radiation detector that constitutes the PET apparatus. An example of an output device connected to the input / output interface 14 is a display. For example, if the information processing device 1 has a function to generate a PET image from a crystal address obtained by crystal identification, the display is used to display this PET image.

[0024] Examples of interfaces that can be used as the input / output interface 14 include PCI (Peripheral Component Interconnect) interfaces and USB (Universal Serial Bus).

[0025] Furthermore, information processing programs P1, P2, and model μ can each be recorded on a computer-readable, non-temporary, tangible recording medium. This recording medium may be the secondary memory 13 or other recording media. For example, tapes, disks, cards, semiconductor memory, programmable logic circuits, etc., can be used as other recording media.

[0026] Furthermore, the information processing device 1 may perform both the information processing method S1 in the inference phase and the information processing method S2 in the learning phase, or it may perform only the information processing method S1 in the inference phase, or it may perform only the information processing method S2 in the learning phase.

[0027] (Flow of Information Processing Method in the Inference Phase) The flow of the information processing method S1 performed by the information processing device 1 in the inference phase will be explained with reference to Figure 2. Figure 2 is a flowchart showing the flow of the information processing method S1.

[0028] The information processing method S1 is a method for performing crystal identification using a model μ generated by machine learning, and as shown in Figure 2, it includes a signal acquisition process S11, a first event identification process S12, a crystal identification process S13, and a second event identification process S14. The processor 11 of the information processing device 1 executes each of the processes included in the information processing method S1 targeting the radiation detector where the event occurred, either whenever an event occurs in any of the multiple radiation detectors constituting the PET device, or all at once after data (signal) collection.

[0029] The signal acquisition process S11 is a process for acquiring the output signal of the light receiving element array included in the target radiation detector. The output signal of the light receiving element array is a set of fluorescence intensities detected by each light receiving element constituting the light receiving element array. In the signal acquisition process S11, the processor 11 acquires the output signal of the light receiving element array included in the target radiation detector from the target radiation detector.

[0030] The first event identification process S12 is a process for identifying whether the event that occurred in the target radiation detector is photoelectric absorption or in-crystal scattering by referring to the output signal of the light receiving element array acquired in the signal acquisition process S11.

[0031] In the present embodiment, since pLUTs optimized for each of the plurality of radiation detectors are not used, the LCR (Light Collection ratio) is used in the first event identification process S12. The LCR is defined as LCR = {max ｈ，ｉ (I １≦ｈ，ｉ≦Ｍ )} / {Σ ｈ，ｉ (I １≦ｈ，ｉ≦Ｍ )}, where I ｈ，ｉ is the fluorescence intensity detected by the light receiving element arranged in the h-th row and i-th column in the light receiving element array composed of M × M light receiving elements. When one light receiving element detects fluorescence and the other M × M - 1 light receiving elements do not detect fluorescence, the LCR takes the maximum value of 1. Conversely, when all M × M light receiving elements detect fluorescence evenly, the LCR takes the minimum value of 1 / (M × M).

[0032] In the first event identification process S12, the processor 11 calculates the LCR by referring to the output signal of the light receiving element array acquired in the signal acquisition process S11, and compares the calculated LCR with a predetermined threshold Th1 (for example, 0.45). When the calculated LCR exceeds the threshold Th1 (when the fluorescence intensity is concentrated), it is identified that the event occurring in the target radiation detector is a photoelectric absorption. On the other hand, when the calculated LCR is lower than the threshold Th1 (when the fluorescence intensity is dispersed), it is identified that the event occurring in the target radiation detector is an in-crystal scattering. Note that although event identification by LCR is a simple method that does not use a pLUT, when the location where Compton scattering occurs and the crystal where photoelectric absorption occurs are close, the output signals of the light receiving element array overlap, making it difficult to identify the event.

[0033] When it is identified in the first event identification process S12 that the event occurring in the target radiation detector is a photoelectric absorption, the processor 11 executes the crystal identification process S13.

[0034] The crystal identification process S13 is a process for identifying the crystal in which photoelectric absorption has occurred in the target radiation detector by referring to the output signal of the light receiving element array acquired in the signal acquisition process S11.

[0035] In the present embodiment, in the crystal identification process S13, the model μ generated by machine learning is used. The model μ takes the output signal of the light receiving element array included in the radiation detector as an input, and outputs a set of likelihoods corresponding to each crystal constituting the crystal array included in the radiation detector, that is, the likelihood that the crystal is the crystal where photoelectric absorption has occurred. The model μ is single. That is, the processor 11 executes the crystal identification process S13 using the same model μ regardless of the individual differences in performance among a plurality of target radiation detectors. A specific example of the model μ will be described later.

[0036] In the crystal identification process S13, the processor 11 inputs the output signal of the photodetector array acquired in the signal acquisition process S11 to the model μ. Then, among the crystals constituting the crystal array included in the target radiation detector, the crystal with the highest likelihood output from the model μ is identified as the crystal where photoelectric absorption occurred. Hereinafter, the highest likelihood output from the model μ, that is, the likelihood corresponding to the crystal where photoelectric absorption occurred, will also be referred to as the reliability of the crystal identification process S13, or the reliability of crystal identification using the model μ.

[0037] The second event identification process S14 is a process that, in order to supplement the identification performance of the first event identification process S12 by LCR described above, refers to the reliability of the crystal identification process S13 to identify whether an event occurring in the target radiation detector is photoelectric absorption or intracrystalline scattering. In the second event identification process S14, the processor 11 compares the reliability of the crystal identification process S13 with a predetermined threshold Th2 (for example, 0.9). If the reliability of the crystal identification process S13 exceeds the threshold Th2, it identifies the event occurring in the target radiation detector as photoelectric absorption. On the other hand, if the reliability of the crystal identification process S13 falls below the threshold Th2, it identifies the event occurring in the target radiation detector as intracrystalline scattering.

[0038] By performing the information processing method S1 on the radiation detector where an event occurred, (1) event identification can be performed to identify whether the event is photoelectric absorption or scattering within a crystal, and (2) if the event is photoelectric absorption, crystal identification can be performed to identify the crystal in which the photoelectric absorption occurred in that radiation detector.

[0039] As described later, Model μ is a general-purpose model obtained by machine learning the relationship between the output signal of the photodetector array and the results of crystal identification using a pLUT for photoelectric absorption occurring at different times in different radiation detectors. Therefore, accurate event identification and crystal identification can be performed without creating a separate Model μ for each radiation detector or periodically updating Model μ.

[0040] In this embodiment, a configuration in which both the first event identification process S12 and the second event identification process S14 are performed has been described, but the present invention is not limited thereto. That is, if the required level for crystal identification is low, either the first event identification process S12 or the second event identification process S14 may be omitted. Furthermore, if the required level for accuracy in crystal identification is even lower and event identification is unnecessary, both the first event identification process S12 and the second event identification process may be omitted.

[0041] Furthermore, although this embodiment describes a configuration in which the likelihood corresponding to each crystal (the likelihood that the crystal is one in which photoelectric absorption occurred) is used as the output of model μ, the present invention is not limited thereto. For example, if information regarding the detection depth of the crystal can be obtained by some method, a configuration may be adopted in which each crystal is divided into slices of different depths, and the likelihood corresponding to each slice of each crystal (the likelihood that the slice is one in which photoelectric absorption occurred) is used as the output of model μ. This makes it possible in the crystal identification process S13 to identify not only which crystals have experienced photoelectric absorption, but also which slices (i.e., at what depth) have experienced photoelectric absorption.

[0042] (Specific Example of the Model) A specific example of the model μ used by the information processing device 1 will be explained with reference to Figure 3. Figure 3 is a schematic diagram showing the configuration of the model μ.

[0043] Model μ is a model generated by machine learning. In Figure 3, a CNN (Convolutional Neural Network) is shown as an example of Model μ.

[0044] The input to Model μ is the output signal of the photodetector array included in the radiation detector, that is, the set of fluorescence intensities corresponding to each photodetector constituting the photodetector array included in the radiation detector. If the photodetector array is composed of M × M photodetectors arranged in a square matrix, the input to Model μ is the M × M fluorescence intensities I ｈ，ｉ This is a set (where h and i are natural numbers between 1 and M). Here, fluorescence intensity I ｈ，ｉThis is the fluorescence intensity detected by the photodetectors arranged in row h and column i of the photodetector array. In Figure 3, the input to model μ is the fluorescence intensity I corresponding to each of the 8x8 photodetectors constituting the photodetector array. ｈ，ｉ This is illustrated as an 8x8 pixel image.

[0045] The output of model μ is the set of likelihoods corresponding to each crystal that makes up the crystal array contained in the radiation detector. If the crystal array consists of N × N crystals arranged in a square matrix, the output of model μ is the set of N × N likelihoods L ｊ，ｋ This is the set (where j and k are natural numbers between 1 and N, inclusive). Here, the likelihood L ｊ，ｋ This is the likelihood that the crystals arranged in the j row and k column of the crystal array are the crystals that exhibited photoelectric absorption. In Figure 3, the output of model μ is the likelihood L corresponding to each of the 14 × 14 crystals that make up the crystal array. ｊ，ｋ This is illustrated as a 14x14 pixel image.

[0046] (Information Processing Flow in the Learning Phase) The flow of the information processing method S2 performed by the information processing device 1 in the learning phase will be explained with reference to Figure 4. Figure 4 is a flowchart showing the flow of the information processing method S2.

[0047] The information processing method S2 is a method for generating a model μ for crystal identification using machine learning (supervised learning in this embodiment), and as shown in Figure 4, it includes a data collection process S21, a centroid calculation process S22, an event identification process S23, a crystal identification process S24, a training data generation process S25, and a model learning process S26.

[0048] The data acquisition process S21 is a process for acquiring data of the output signal of the photodetector array contained in the target radiation detector. The output signal of the photodetector array is a collection of fluorescence intensities detected by each photodetector constituting the photodetector array. In the data acquisition process S21, data of the output signal of the photodetector array contained in the target radiation detector is acquired. The centroid calculation process S22, event identification process S23, crystal identification process S24, training data generation process S25, and model learning process S26 are executed using the output signal data acquired in the data acquisition process S21.

[0049] The centroid calculation process S22 is a process that calculates the centroid position by referring to the output signal data of the photodetector array collected in the data acquisition process S21. In the centroid calculation process S22, the processor 11 sets the coordinates (X, Y) of the centroid position as X = (Σ １≦ｈ，ｉ≦Ｍ I ｈ，ｉ ×x ｈ，ｉ ) / (Σ １≦ｈ，ｉ≦Ｍ I ｈ，ｉ ), Y = (Σ １≦ｈ，ｉ≦Ｍ I ｈ，ｉ ×y ｈ，ｉ ) / (Σ １≦ｈ，ｉ≦Ｍ I ｈ，ｉ It is calculated according to ). Here, I ｈ，ｉ This is the fluorescence intensity detected by the photodetector located in row h, column i, and (x ｈ，ｉ , y ｈ，ｉ ) is the center coordinate of the light-receiving elements arranged in h rows and i columns.

[0050] The event identification process S23 is a process that refers to the centroid position calculated in the centroid calculation process S22 to identify whether the event that occurred in the target radiation detector is photoelectric absorption or in-crystal scattering.

[0051] In this embodiment, in the event identification process S23, a pLUT calibrated specifically for the target radiation detector is used. The pLUT is a table that associates the address of the crystal where photoelectric absorption occurred with the centroid position obtained by centroid calculation. In the event identification process S23, the processor 11 identifies that the event that occurred in the target radiation detector is photoelectric absorption if the centroid position calculated in the centroid calculation process S22 is associated with any crystal (address) in the pLUT. On the other hand, if the centroid position calculated in the centroid calculation process S22 is not associated with any crystal (address) in the pLUT, the processor 11 identifies that the event that occurred in the target radiation detector is intracrystal scattering.

[0052] The processor 11 uses only the events identified in the event identification process S23 as photoelectric absorption events that occurred in the target radiation detector to execute the crystal identification process S24 and the training data generation process S25.

[0053] The crystal identification process S24 is a process for identifying the crystal that has produced photoelectric absorption in the target radiation detector by referring to the centroid position calculated in the centroid calculation process S22.

[0054] In this embodiment, in the crystal identification process S24, a pLUT calibrated specifically for the target radiation detector is used. As described above, the pLUT is a table that associates the address of the crystal where photoelectric absorption occurred with the centroid position obtained by centroid calculation. In the crystal identification process S24, the processor 11 uses the address of the crystal associated in the pLUT with the centroid position calculated in the centroid calculation process S22 as the address of the crystal where photoelectric absorption occurred in the target radiation detector.

[0055] The training data generation process S25 is a process for generating training data to be used for machine learning of model μ by referring to the output signals of the photodetector array acquired in the data acquisition process S21 and the addresses of the crystals identified in the crystal identification process S24. In the training data generation process S25, the processor 11 generates training data by adding the addresses of the crystals identified in the crystal identification process S24 as correct labels to the output signals of the photodetector array acquired in the data acquisition process S21.

[0056] The model learning process S26 is a process for learning the model μ by referring to the training data generated in the training data generation process S25. In the model learning process S26, the processor 11 learns the parameters that constitute the model μ such that (1) the likelihood corresponding to the crystal whose address is included as the correct label in the training data (the crystal that produced photoelectric absorption) is high, and (2) the likelihood corresponding to other crystals (crystals other than the crystal that produced photoelectric absorption) is low, when the output signal of the photodetector array included in the training data is input to the model μ.

[0057] By performing the information processing method S2 on the radiation detector where the event occurred, either each time an event occurs in any of the multiple radiation detectors constituting the PET apparatus, or collectively after data (signal) collection, the model μ can learn the relationship between the output signal of the photodetector array and the crystal identification result using the pLUT for photoelectric absorption that occurred at different times in different radiation detectors. This makes it possible to generate a general-purpose model μ that does not need to be created for each radiation detector or updated periodically.

[0058] (Example) The above-described information processing method S2 was carried out using a PET device for small animals equipped with 126 radiation detectors. Each of the 126 radiation detectors is equipped with a Fast-LGSO scintillator array consisting of 14 × 14 crystals and a multi-pixel photon counter (MPPC) consisting of 8 × 8 photodetectors. A CNN was used as the model μ.

[0059] First, 40 radiation detectors were selected from the 126 radiation detectors that make up the PET apparatus. Then, machine learning (supervised learning) of model μ was performed by applying information processing method S2 to each of the 5 million events generated by these 40 radiation detectors. Subsequently, model μ was evaluated using the 86 radiation detectors excluding the 40 used for machine learning. As a result of the evaluation, the accuracy rate of crystal identification using model μ was 96%. Note that the accuracy rate of crystal identification using model μ refers to the probability that the crystal identification result using model μ and the crystal identification result using pLUT match in a photoelectric absorption event.

[0060] Figure 5 shows images representing the input to Model μ (the set of fluorescence intensities detected by each photodetector) and the output to Model μ (the set of likelihoods that each crystal is a crystal that has undergone photoelectric absorption) for each of the 20 events in which the crystal identification result was correct and the 10 events in which the crystal identification result was incorrect. The natural numbers shown in the upper left / upper right of the image representing the input to Model μ are the results of crystal identification using pLUT / Model μ (the crystal address), respectively. For the 20 events in which the crystal identification result was correct, it can be confirmed from Figure 5 that the results of crystal identification using pLUT and the results of crystal identification using Model μ are consistent. On the other hand, for the 10 events in which the crystal identification result was incorrect, it can be confirmed from Figure 5 that the results of crystal identification using pLUT and the results of crystal identification using Model μ are not consistent.

[0061] Furthermore, the majority of events in which the crystal identification result was incorrect were events that were judged as photoelectric absorption by the pLUT crystal identification result, but were actually intracrystal scattering. In crystal identification by pLUT, the photoelectric absorption region (A ｉ，ｊ Because intracrystalline scattering that overlaps with photoelectric absorption cannot be distinguished from photoelectric absorption, misidentification is likely to occur.

[0062] Next, a phantom was placed in the center of the field of view of the PET device for small animals described above, and the information processing method S1 according to the example and the information processing method according to the comparative example were performed. Here, the information processing method according to the comparative example refers to a method in which, in information processing method S1, (1) the first event identification process S12 (event identification using LCR) is replaced with event identification using pLUT, (2) the crystal identification process S13 (crystal identification using model μ) is replaced with crystal identification using pLUT, and (3) the second event identification process S14 (event identification using the confidence level of crystal identification process S13) is omitted. Then, a timing histogram and a reconstructed image were generated by referring to the crystal address obtained by the information processing method S1 according to the example. Similarly, a timing histogram and a reconstructed image were generated by referring to the crystal address obtained by the information processing method according to the comparative example.

[0063] Figure 6(a) is a timing histogram generated by referencing the crystal address obtained by the information processing method according to the comparative example, and Figure 6(b) is a timing histogram generated by referencing the crystal address obtained by the information processing method S1 according to the example. Intracrystal scattering, where energy is dispersed among multiple crystals, degrades the time resolution because the fluorescence output at each detection element becomes low. Comparing these two timing histograms, it can be seen that (1) the time resolution in information processing method S1 according to the example is equivalent to the time resolution in the information processing method according to the comparative example, and (2) the sensitivity in information processing method S1 according to the example is higher than the sensitivity in the information processing method according to the comparative example. In the comparative example, high time resolution was obtained by reducing intracrystal scattering as much as possible using a pLUT, but sensitivity was sacrificed. On the other hand, in the example, high sensitivity and time resolution can be achieved simultaneously by incorporating intracrystal scattering, which has little energy transfer by Compton scattering and has little impact on crystal identification.

[0064] Figure 7(a) is a reconstructed image of a rod-shaped phantom with multiple diameters generated by referencing the crystal address obtained by the information processing method according to the comparative example, and Figure 7(b) is a reconstructed image of a rod-shaped phantom with multiple diameters generated by referencing the crystal address obtained by the information processing method S1 according to the example. Comparing these two reconstructed images, it can be seen that the spatial resolution in the information processing method S1 according to the example is equivalent to the spatial resolution in the information processing method according to the comparative example.

[0065] In the comparative example's information processing method, 128 pLUTs were used, whereas in the example's information processing method S1, only one model μ was used. Furthermore, 40 pLUTs were used in information processing method S2 to generate this model μ. Despite this, the temporal and spatial resolution when implementing the example's information processing method S1 was comparable to that when implementing the comparative example's information processing method.

[0066] (Summary) The information processing device according to Embodiment 1 is characterized in that it comprises a processor, the processor performs a crystal identification process that identifies the crystal in which photoelectric absorption occurred in a radiation detector where an event occurred, from among a plurality of radiation detectors, each including a crystal array and a photodetector array, using a single model generated by machine learning, the model takes the output signal of the photodetector array included in the radiation detector as input and outputs the likelihood corresponding to each crystal constituting the crystal array included in the radiation detector, which is the likelihood that the crystal is a crystal in which photoelectric absorption occurred.

[0067] According to the above configuration, crystal identification in multiple radiation detectors can be performed without using pLUTs. Therefore, since there is no need to create a pLUT for each radiation detector or to periodically update the pLUTs, the manufacturing and operating costs of equipment including multiple radiation detectors (e.g., PET scanners) can be reduced.

[0068] The information processing device according to Embodiment 2 is characterized in that, in the information processing device according to Embodiment 1, the processor calculates an LCR (Light Collection Ratio) by referring to the output signal of a photodetector array included in a radiation detector, and further performs a first event identification process to identify whether the event is photoelectric absorption or in-crystal scattering by comparing the calculated LCR with a first threshold, and if the processor identifies the event as photoelectric absorption in the first event identification process, it performs the crystal identification process.

[0069] With the above configuration, crystal identification can be performed with high accuracy without using a pLUT for events identified as photoelectric absorption in the first event identification process.

[0070] The information processing device according to embodiment 3 is characterized in that, in the information processing device according to embodiment 1 or 2, the processor further performs a second event identification process in which it identifies whether the event is photoelectric absorption or intracrystal scattering by comparing the highest likelihood among the likelihoods output from the model in the crystal identification process with a second threshold.

[0071] According to the above configuration, events with low reliability in the crystal identification process are identified as intracrystal scattering in the second event identification process. Therefore, crystal identification can be performed with high accuracy for events identified as photoelectric absorption in the second event identification process.

[0072] The information processing device according to embodiment 4 includes a processor, the processor performs a centroid calculation process on the output signal of the photodetector array included in a plurality of radiation detectors, each including a crystal array and a photodetector array, where the event occurred; an event identification process that identifies whether the event is photoelectric absorption or in-crystal scattering by comparing the centroid position calculated in the centroid calculation process with the pLUT; and if the event is identified as photoelectric absorption in the event identification process, the processor determines whether the event occurred by comparing the centroid position calculated in the centroid calculation process with the pLUT. The method is characterized by performing: a crystal identification process to identify a crystal in which photoelectric absorption has occurred in a radiation detector; a training data generation process to generate training data by adding the address of the crystal identified in the crystal identification process as a correct label to the output signal of the photodetector array included in the radiation detector in which the event occurred; and a model learning process that uses the training data generated in the training data generation process to learn a model that takes the output signal of the photodetector array included in the radiation detector as input and outputs the likelihood corresponding to each crystal constituting the crystal array included in the radiation detector, where the crystal is the crystal in which photoelectric absorption has occurred.

[0073] According to the above configuration, a model can be generated that takes the output signal of the photodetector array contained in the radiation detector as input and outputs the likelihood of each crystal constituting the crystal array contained in the radiation detector, where the crystal is the crystal that has undergone photoelectric absorption. Using this model, crystal identification in multiple radiation detectors can be performed without using pLUTs. Although pLUTs are used to create training data, only a small number of radiation detectors' pLUTs are needed, making the creation of training data easy.

[0074] The information processing method according to embodiment 5 includes a crystal identification process in which a processor identifies a crystal that has undergone photoelectric absorption in a radiation detector where an event has occurred, using a single model generated by machine learning, among a plurality of radiation detectors, each of which includes a crystal array and a photodetector array, wherein the model takes the output signal of the photodetector array included in the radiation detector as input and outputs the likelihood corresponding to each crystal constituting the crystal array included in the radiation detector, which is the likelihood that the crystal has undergone photoelectric absorption.

[0075] According to the above configuration, crystal identification in multiple radiation detectors can be performed without using pLUTs. Therefore, since there is no need to create a pLUT for each radiation detector or to periodically update the pLUTs, the manufacturing and operating costs of equipment including multiple radiation detectors (e.g., PET scanners) can be reduced.

[0076] The information processing method according to embodiment 6 includes a centroid calculation process in which a processor performs a centroid calculation on the output signal of the photodetector array included in a plurality of radiation detectors, each of which includes a crystal array and a photodetector array, where an event occurred, and the processor performs a centroid calculation on the centroid position calculated in the centroid calculation process and pLUT (position Look Up The invention is characterized by comprising: an event identification process that identifies whether the event is photoelectric absorption or in-crystal scattering by comparing it with a Table; a crystal identification process in which, if the event is identified as photoelectric absorption in the event identification process, the processor identifies the crystal in which photoelectric absorption occurred in the radiation detector where the event occurred by comparing the centroid position calculated in the centroid calculation process with the pLUT; a training data generation process in which the processor generates training data by adding the address of the crystal identified in the crystal identification process as a correct label to the output signal of the photodetector array included in the radiation detector where the event occurred; and a model learning process in which the processor learns a model that takes the output signal of the photodetector array included in the radiation detector as input and outputs the likelihood corresponding to each crystal constituting the crystal array included in the radiation detector, wherein the crystal is the crystal in which photoelectric absorption occurred, using the training data generated in the training data generation process.

[0077] According to the above configuration, a model can be generated that takes the output signal of the photodetector array contained in the radiation detector as input and outputs the likelihood corresponding to each crystal constituting the crystal array contained in the radiation detector, where the crystal is the crystal that has undergone photoelectric absorption. Furthermore, using this model, crystal identification in multiple radiation detectors can be performed without using a pLUT.

[0078] The PET (Positron Emission Tomography) apparatus according to embodiment 7 is characterized by comprising an information processing device according to any one embodiment of embodiments 1 to 3, and the plurality of radiation detectors.

[0079] According to the above configuration, crystal identification in the multiple radiation detectors constituting the PET apparatus can be performed without using pLUTs. Therefore, since there is no need to create a pLUT for each radiation detector or to periodically update the pLUTs, the manufacturing and operating costs of the PET can be reduced.

[0080] The program according to embodiment 8 is a program that causes a computer equipped with a processor to execute the information processing method described in claim 5, characterized in that it causes the processor to execute the crystal identification process.

[0081] According to the above configuration, crystal identification in the multiple radiation detectors constituting the PET apparatus can be performed without using pLUTs. Therefore, since there is no need to create a pLUT for each radiation detector or to periodically update the pLUTs, the manufacturing and operating costs of the PET can be reduced.

[0082] The program according to embodiment 9 is a program that causes a computer equipped with a processor to execute the information processing method described in claim 6, characterized in that it causes the processor to execute the centroid calculation process, the event identification process, the crystal identification process, the training data generation process, and the model learning process.

[0083] The computer-readable recording medium according to embodiment 10 is characterized in that it records the program described in embodiment 8 or 9.

[0084] According to the above configuration, crystal identification in the multiple radiation detectors constituting the PET apparatus can be performed without using pLUTs. Therefore, since there is no need to create a pLUT for each radiation detector or to periodically update the pLUTs, the manufacturing and operating costs of the PET can be reduced.

[0085] (Additional Notes) The present invention is not limited to the embodiments described above, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in the embodiments described above are also included in the technical scope of the present invention.

[0086] 1 Information Processing Device 11 Processor 12 Primary Memory 13 Secondary Memory 14 Input / Output Interface 15 Bus μ Model S1 Information Processing Method (Inference Phase) S11 Signal Acquisition Process S12 First Event Identification Process S13 Crystal Identification Process S14 Second Event Identification Process S2 Information Processing Method (Learning Phase) S21 Data Acquisition Process S22 Centroid Calculation Process S23 Event Identification Process S24 Crystal Identification Process S25 Training Data Generation Process S26 Model Learning Process

Claims

1. An information processing device comprising a processor, the processor performing a crystal identification process to identify, using a single model generated by machine learning, the crystal in which photoelectric absorption occurred in a radiation detector where an event occurred, from among a plurality of radiation detectors, each including a crystal array and a photodetector array, the model taking the output signal of the photodetector array included in the radiation detector as input, and outputting the likelihood corresponding to each crystal constituting the crystal array included in the radiation detector, which is the likelihood that the crystal is a crystal in which photoelectric absorption occurred.

2. The information processing apparatus according to claim 1, characterized in that the processor calculates a Light Collection Ratio (LCR) by referring to the output signal of a photodetector array included in a radiation detector, and further performs a first event identification process to identify whether the event is photoelectric absorption or in-crystal scattering by comparing the calculated LCR with a first threshold, and if the processor identifies the event as photoelectric absorption in the first event identification process, it performs the crystal identification process.

3. The information processing apparatus according to claim 1 or 2, characterized in that the processor further performs a second event identification process in which it identifies whether the event is photoelectric absorption or intracrystal scattering by comparing the highest likelihood among the likelihoods output from the model in the crystal identification process with a second threshold.

4. A processor comprising: a plurality of radiation detectors, each including a crystal array and a photodetector array, performing a centroid calculation on the output signal of the photodetector array included in the radiation detector where the event occurred; an event identification process that identifies whether the event is photoelectric absorption or in-crystal scattering by comparing the centroid position calculated in the centroid calculation process with the pLUT; a crystal identification process that, if the event is identified as photoelectric absorption in the event identification process, identifies the crystal in which photoelectric absorption occurred in the radiation detector where the event occurred by comparing the centroid position calculated in the centroid calculation process with the pLUT; and a training data generation process that generates training data by adding the address of the crystal identified in the crystal identification process as a correct label to the output signal of the photodetector array included in the radiation detector where the event occurred. An information processing device characterized by performing a model learning process that uses the training data generated in the training data generation process to learn a model that takes the output signal of a photodetector array contained in a radiation detector as input and outputs the likelihood corresponding to each crystal constituting the crystal array contained in the radiation detector, wherein the crystal is a crystal that has undergone photoelectric absorption.

5. An information processing method characterized in that the processor includes a crystal identification process in which, among a plurality of radiation detectors, each including a crystal array and a photodetector array, a crystal in which photoelectric absorption occurred in a radiation detector where an event occurred is identified using a single model generated by machine learning, wherein the model takes the output signal of the photodetector array included in the radiation detector as input and outputs the likelihood corresponding to each crystal constituting the crystal array included in the radiation detector, which is the likelihood that the crystal is a crystal in which photoelectric absorption occurred.

6. A centroid calculation process in which the processor performs a centroid calculation on the output signal of the photodetector array included in a plurality of radiation detectors, each including a crystal array and a photodetector array, where the event occurred; an event identification process in which the processor identifies whether the event is photoelectric absorption or intracrystal scattering by comparing the centroid position calculated in the centroid calculation process with a pLUT (position look-up table); if the event is identified as photoelectric absorption in the event identification process, a crystal identification process in which the processor identifies the crystal in which photoelectric absorption occurred in the radiation detector where the event occurred by comparing the centroid position calculated in the centroid calculation process with the pLUT; and a training data generation process in which the processor generates training data by adding the address of the crystal identified in the crystal identification process as a correct label to the output signal of the photodetector array included in the radiation detector where the event occurred. An information processing method characterized in that the processor includes a model learning process which uses the training data generated in the training data generation process to learn a model that takes the output signal of a photodetector array included in a radiation detector as input and outputs the likelihood corresponding to each crystal constituting the crystal array included in the radiation detector, wherein the crystal is a crystal that has undergone photoelectric absorption.

7. A PET (Positron Emission Tomography) apparatus comprising: an information processing device according to any one of claims 1 to 3; and the plurality of radiation detectors.

8. A program that causes a computer equipped with a processor to execute the information processing method described in claim 5, wherein the program causes the processor to execute the crystal identification process.

9. A program that causes a computer equipped with a processor to execute the information processing method described in claim 6, wherein the program causes the processor to execute the centroid calculation process, the event identification process, the crystal identification process, the training data generation process, and the model learning process.

10. A computer-readable recording medium having the program described in claim 8 or 9 recorded on it.

Images