JP7395269B2 - Medical information processing equipment - Google Patents

Description

Embodiments of the present invention relate to a medical information processing device.

Nuclear medicine diagnostic equipment such as PET (Positron Emission Tomography) is equipped with a gamma ray detector including multiple scintillator crystals and multiple optical sensors. The scintillator crystal converts the incident gamma rays into fluorescence, and the optical sensor detects the fluorescence. The scintillator crystal into which gamma rays have been input is identified by comparing the coordinates obtained by calculating the center of gravity of the output signal from the optical sensor with a look-up table (LUT). The lookup table is generated by dividing a flood map representing the spatial distribution of the frequency of fluorescence originating from gamma rays into a plurality of regions and associating each region with a scintillator crystal. The divisions are generated based on the flood map using a predetermined algorithm. However, the algorithm is not perfect and requires manual modification. Modification is work that requires man-hours. The same type of modification work is repeated until the algorithm is updated. This is inefficient. Furthermore, since the performance of gamma ray detectors changes over time, it is also necessary to perform periodic classification work.

US Patent Application Publication No. US2015/0276953 US Patent No. 9672617 US Patent No. 9668699

The problem to be solved by the present invention is to easily and accurately divide a flood map.

The medical information processing device according to the embodiment includes an acquisition unit that acquires a flood map to be processed that indicates the spatial distribution of the frequency of fluorescence originating from gamma rays, and a flood map that indicates the spatial distribution of the frequency of fluorescence originating from gamma rays. a machine learning model that outputs a division in a flood map for identifying the scintillator crystal in which the fluorescence has occurred based on the processing target; and a division generation unit that generates divisions in the flood map to be processed based on the flood map to be processed. and.

FIG. 1 is a diagram showing the configuration of a PET apparatus according to this embodiment. FIG. 2 is a side view of the gamma ray detector of FIG. 1. FIG. 3 is a diagram showing a plan view and a flood map of the gamma ray detector of FIG. 1. FIG. 4 is a diagram showing the flood map and crystal divisions of FIG. 3. FIG. 5 is a diagram showing an example of the center of gravity/crystal table of FIG. 1. FIG. 6 is a diagram showing a typical flow of a series of processes related to generation of a center of gravity/crystal table by the processing circuit according to the present embodiment. FIG. 7 is a diagram schematically showing a typical flow of the process in FIG. 6. FIG. 8 is a diagram showing an example of a display screen that is displayed in step SA3 of FIG. 6 and includes a GUI that accepts instructions for modifying crystal classification. FIG. 9 is a diagram showing an example of the modification of crystal divisions performed in step SA4 of FIG. 6. FIG. 10 is a diagram showing a typical flow of a series of processes related to generation of a center of gravity/crystal table by a processing circuit according to a modification of the present embodiment. FIG. 11 is a diagram schematically showing a typical flow of the process in FIG. 10. FIG. 12 is a diagram showing the configuration of a medical image processing apparatus according to another example of this embodiment. FIG. 13 is a diagram schematically showing a typical flow of processing for generating crystal sections from a flood map by the processing circuit according to Application Example 1. FIG. 14 is a diagram showing the relationship between a flood map and a divided flood map. FIG. 15 is a diagram schematically showing the integration process performed in step SC2 of FIG. 13. FIG. 16 is a diagram schematically showing a typical flow of a process of generating crystal sections from a flood map by a processing circuit according to Application Example 2. FIG. 17 is a diagram illustrating an example of a display screen including a GUI that accepts an instruction to designate a patch application position. FIG. 18 is a diagram showing the relationship between the cut-out crystal sections of the patch size and the temporary crystal sections of the existing size, which are integrated in the integration process of FIG. 16.

The medical information processing device according to the present embodiment is a computer for at least dividing scintillator crystals in a flood map that expresses the spatial distribution of the frequency of fluorescence derived from gamma rays. This classification work involves associating the center of gravity calculation results with the incident scintillator crystal, which is used in nuclear medicine diagnostic equipment such as PET (Positron Emission Tomography), SPECT (Single Photon Emission CT), PET/CT equipment, and SPECT/CT equipment. done to generate data. The medical information processing device according to this embodiment may be incorporated into the above-mentioned nuclear medicine diagnostic device or may be a separate computer, but in order to specifically explain the following, it is incorporated into the PET device. It is assumed that

FIG. 1 is a diagram showing the configuration of a PET apparatus 1 according to this embodiment. As shown in FIG. 1, the PET apparatus 1 includes a gantry 10, a bed 30, and a console 50. The gantry 10 is an imaging device that performs PET imaging of the subject P. The subject P is placed on the bed 30. The console 50 is a computer that controls the gantry 10 and the bed 30. The console 50 is a type of medical information processing device according to this embodiment.

As shown in FIG. 1, the gantry 10 includes a detector ring 11, a detection processor 13, and a coincidence circuit 15.

The detector ring 11 has a plurality of gamma ray detectors 17 arranged on the circumference around the central axis Z. An image field of view (FOV) is set in the opening of the detector ring 11. The subject P is positioned so that the imaged region of the subject P is included in the image field of view. A drug labeled with a positron-emitting nuclide is administered to the subject P. Positrons emitted from positron-emitting nuclides are annihilated with surrounding electrons, producing a pair of annihilation gamma rays. The gamma ray detector 17 detects gamma rays emitted from the body of the subject P, and generates an electrical signal according to the amount of detected annihilation gamma rays. Gamma ray detector 17 includes multiple scintillator crystals and multiple optical sensors. The scintillator crystal interacts with annihilation gamma rays originating from radioactive isotopes within the subject P, and generates a plurality of fluorescent photons. The interaction event between the scintillator crystal and gamma rays is also called a single event. The optical sensor receives fluorescent photons and generates an electrical signal according to the amount of the received fluorescent photons. As the optical sensor, a photomultiplier tube, a photodiode, a semiconductor detection element, etc. are used. The generated electrical signal is supplied to the connected detection processor 13.

Here, the central axis of the detector ring 11 is defined as the Z-axis, the axis perpendicular to the floor surface as the Y-axis, and the axis perpendicular to the Z-axis and the Y-axis as the X-axis. The XYZ coordinate system forms a three-dimensional orthogonal coordinate system.

The plurality of detection processors 13 are connected to the plurality of gamma ray detectors 17, respectively. The detection processor 13 generates single event data based on the electrical signal from the gamma ray detector 17 to which it is connected. Specifically, the detection processor 13 includes a processing circuit 19 and a memory 21. The processing circuit 19 includes a processor such as a CPU (Central Processing Unit). Specifically, the processing circuit 19 realizes a center of gravity calculation function 23 and an energy calculation function 25.

In the center of gravity calculation function 23, the processing circuit 19 calculates the incident position of gamma rays based on the electrical signal from the connected gamma ray detector 17. The incident position of the gamma ray is represented by the scintillator crystal into which the gamma ray is incident. Specifically, the processing circuit 19 performs a center of gravity calculation on the peak values of a plurality of output signals from a plurality of optical sensors, and calculates the position coordinates (center of gravity) of a gamma ray interaction event occurring in a scintillator crystal. . Next, the processing circuit 19 checks the calculated center of gravity against the center of gravity/crystal table 27 stored in the memory 21 to identify the scintillator crystal in which the gamma ray interaction event has occurred.

In the energy calculation function 25, the processing circuit 19 calculates the energy value of the gamma ray that caused the interaction event in the gamma ray detector 17 that is the source, based on the electrical signal from the gamma ray detector 17 that is the source. Specifically, the processing circuit 19 generates a sum signal of a plurality of output signals from a plurality of optical sensors, and specifies the peak value of the sum signal as energy.

The incident scintillator crystal identifier and energy value for each interaction event are associated with the detection time. The detection time is measured by a time digital converter or the like included in the processing circuit 19. For example, the processing circuit 19 monitors the peak value of a summed signal of a plurality of output signals from a plurality of optical sensors, and measures the time when the peak value exceeds a preset threshold value as the detection time. The combination of the incident scintillator crystal identifier, energy value, and detection time is called single event data. Single event data is generated one after another each time a gamma ray is detected. The generated single event data is supplied to the coincidence counting circuit 15.

The coincidence counting circuit 15 performs coincidence counting processing on the single event data from the plurality of detection processors 13. As a hardware resource, the coincidence circuit 15 includes a processor such as a CPU. The coincidence circuit 15 repeatedly identifies single event data regarding two single events falling within a predetermined time frame from among the single event data that is repeatedly supplied. This pair of single events is estimated to originate from annihilation gamma rays generated from the same annihilation point. Paired single events are collectively referred to as coincidence events. A line connecting the pair of gamma ray detectors 17 (more specifically, scintillator crystals) that detected this annihilation gamma ray is called LOR (line of response). Event data regarding a pair of events forming an LOR is called coincidence event data. The coincidence event data and single event data are transmitted to the console 50. Incidentally, when there is no particular distinction between coincidence event data and single event data, they will be referred to as PET event data.

In the above configuration, the detection processor 13 and the coincidence circuit 15 are included in the gantry 10, but the present embodiment is not limited to this. For example, the coincidence circuit 15 or both the detection processor 13 and the coincidence circuit 15 may be included in a separate device from the gantry 10. Further, the coincidence circuit 15 may be provided one by one for each of the plurality of detection processors 13 mounted on the gantry 10, or the plurality of detection processors 13 mounted on the gantry 10 may be grouped into a plurality of groups. It may be divided and provided one for each group.

As shown in FIG. 1, the console 50 includes a processing circuit 51, a display 52, a memory 53, and an input interface 54. For example, data communication between processing circuit 51, display 52, memory 53, and input interface 54 is performed via a bus.

The processing circuit 51 has a processor such as a CPU or a GPU (Graphics Processing Unit) as a hardware resource. The processing circuit 51 includes a reconstruction function 511, an image processing function 512, an imaging control function 513, a flood map acquisition function 514, a crystal division generation function 515, a crystal division correction function 516, a table generation function 517, a model update function 518, and display control. Function 519 is realized. Note that the reconstruction function 511, image processing function 512, imaging control function 513, flood map acquisition function 514, crystal division generation function 515, crystal division correction function 516, table generation function 517, model update function 518, and display control function 519 are , may be mounted by the processing circuit 51 on one board, or may be mounted in a distributed manner by the processing circuits 51 on a plurality of boards.

In the reconstruction function 511, the processing circuit 51 reconstructs a PET image showing the distribution of the positron-emitting nuclide administered to the subject P based on the coincidence event data transmitted from the gantry 10. As the image reconstruction algorithm, any existing image reconstruction algorithm such as the FBP (filtered back projection) method or the successive approximation reconstruction method may be used.

In the image processing function 512, the processing circuit 51 performs various image processing on the PET image reconstructed by the reconstruction function 511. For example, the processing circuit 51 performs three-dimensional image processing such as volume rendering, surface volume rendering, pixel value projection processing, MPR (Multi-Planer Reconstruction) processing, and CPR (Curved MPR) processing on a PET image to create a display image. generate.

In the imaging control function 513, the processing circuit 51 synchronously controls the gantry 10 and the bed 30 to perform PET imaging. Note that PET imaging for generating the flood map is performed in the absence of the subject P, such as at the time of product shipment or maintenance.

In the flood map acquisition function 514, the processing circuit 51 acquires a flood map that expresses the spatial distribution of the frequency of fluorescence originating from gamma rays. For example, the processing circuit 51 generates a flood map based on the center of gravity calculated by the center of gravity calculation function 23.

Here, the relationship between the gamma ray detector 17 and the flood map will be explained.

FIG. 2 is a side view of the gamma ray detector 17. The side view of FIG. 2 refers to a side view seen from the Z-axis direction. Here, the central axis Z of the gantry 10 is defined as the z-axis, the axis perpendicular to the z-axis is defined as the y-axis, and the axis perpendicular to the z-axis and the y-axis is defined as the x-axis. The xyz three-dimensional coordinate system forms a three-dimensional orthogonal coordinate system with the gamma ray detector 17 as a reference.

As shown in FIG. 2, the gamma ray detector 17 has a plurality of scintillator crystals 61 arranged two-dimensionally about the z-axis and the x-axis. A plurality of optical sensors 63 are arranged two-dimensionally with respect to the z-axis and the x-axis on the back surface of the plurality of scintillator crystals 61 via a light guide 62. Hereinafter, it is assumed that the optical sensor 63 is a photomultiplier tube. The number of photomultiplier tubes 63 provided in each gamma ray detector 17 is smaller than the number of scintillator crystals 61. That is, a plurality of scintillator crystals 61 are connected to the light receiving surface of one photomultiplier tube 63 via a light guide 62.

FIG. 3 is a diagram showing a plan view of the gamma ray detector 17 and a flood map FM1. The upper part of FIG. 3 shows a plan view of the light receiving surface of the gamma ray detector 17, and the lower part of FIG. 3 shows a flood map FM1 regarding the region Z3. As shown in FIG. 3, a plurality of photomultiplier tubes 63 are arranged on the back side of a plurality of scintillator crystals. The plurality of photomultiplier tubes 63 include a photomultiplier tube 63 having a light receiving surface with a relatively small area and a photomultiplier tube 63 having a light receiving surface with a relatively large area. The photomultiplier tubes 63 having a small-area light-receiving surface and the photomultiplier tubes 63 having a large-area light-receiving surface are arranged alternately with respect to the z-axis and the x-axis.

PET imaging for generating a flood map is performed in the absence of the subject P, such as during product shipment or maintenance. In PET imaging for generating a flood map, gamma rays are incident on the gamma ray detector 17 from the gamma ray emitter. Gamma rays incident on a certain scintillator crystal 61 interact with the scintillator crystal 61, and are converted into a plurality of fluorescent photons, the number of which corresponds to the energy of the incident gamma rays. Each fluorescent photon is detected by a photomultiplier tube 63 via a light guide. The detected fluorescent photons are converted by the photomultiplier tube 63 into an electrical signal having a peak value corresponding to the energy of the fluorescent photons. The electrical signal is supplied to a detection processor 13.

The processing circuit 19 of the detection processor 13 receives a plurality of electrical signals from a plurality of connected photomultiplier tubes 63 for each gamma ray interaction event. The processing circuit 19 uses the center of gravity calculation function 23 to perform center of gravity calculation on the plurality of electrical signals. Specifically, the processing circuit 19 calculates the coordinates of the center of gravity based on the peak value of the electrical signal and the position of the photomultiplier tube 63 from which the electrical signal is output in the gamma ray detector coordinate system. Specifically, the gamma ray detector coordinate system is a two-dimensional orthogonal coordinate system defined by the z-axis and x-axis of the gamma ray detector 17. Data on the center of gravity coordinates for each interaction event is supplied to the processing circuit 51.

During PET imaging for generating a flood map, the processing circuit 51 executes a flood map acquisition function 514. The processing circuit 51 plots the barycenter coordinates for each interaction event on the gamma ray detector coordinate system, as shown in the lower part of FIG. As a result, a flood map FM1 representing the spatial distribution of the frequency of fluorescence originating from gamma rays is generated. As shown in the upper part of FIG. 3, the gamma ray detector 17 is divided into a plurality of regions Z1-Z4 along the x-axis. A flood map FM1 is generated for each region Z1-Z4. Note that a flood map covering the entire area Z1-Z4 may be generated.

In the crystal section generation function 515, the processing circuit 51 generates sections (hereinafter referred to as crystals) in the flood map to be processed from the flood map to be processed acquired by the flood map acquisition function 514 according to the machine learning model 531 stored in the memory 53. (referred to as partitions).

FIG. 4 is a diagram showing the flood map FM2 and crystal section SL2 of FIG. 3. As shown in FIG. 4, the crystal section SL2 is line data that indicates a section of the scintillator crystal defined by the coordinate system of the flood map FM2 (ie, the gamma ray detector coordinate system). Due to variations in the light emission characteristics of scintillator crystals and the output characteristics of optical sensors, the interaction event does not actually occur in scintillator crystals located at the center of gravity coordinates, and the interaction event occurs in scintillator crystals in the vicinity. An event may occur. Therefore, it is necessary to associate the barycenter coordinates with the scintillator crystal in which the interaction event corresponding to the barycenter coordinates has occurred, and it is necessary to set crystal classifications in the flood map FM1 as pre-processing for the association.

The machine learning model 531 according to the present embodiment is a composite function with parameters, which is a combination of a plurality of functions, which is trained to input a flood map and output a crystal section corresponding to the flood map. A parameterized composite function is defined by a combination of multiple tunable functions and parameters. Parameter is a general term for weighting matrix and bias. The machine learning model 531 may be any synthetic function with parameters that satisfies the above requirements, such as a deep neural network (DNN).

In the crystal division modification function 516, the processing circuit 51 modifies the crystal division generated in the crystal division generation function 515 according to a user instruction via the input interface 54 or the like.

In the table generation function 517, the processing circuit 51 combines the crystal divisions generated by the crystal division generation function 515 or the crystal divisions after correction by the crystal division correction function 516, and the flood map to be processed acquired by the flood map acquisition function 514. Based on this, association data between a position in the flood map and an identifier of a scintillator crystal corresponding to the position is generated. The association data is realized by a lookup table, a database, or the like. The position in the flood map is equal to the position of the center of gravity calculated by the center of gravity calculation function 23. The scintillator crystal identifier is information such as name, position, number, etc. that can uniquely identify the scintillator crystal. Hereinafter, this association data will be referred to as the center of gravity/crystal table 27. The center of gravity/crystal table 27 is generated for each gamma ray detector 17. The center of gravity/crystal table 27 is stored in the memory 21 of the corresponding detection processor 13.

FIG. 5 is a diagram showing an example of the center of gravity/crystal table 27. As shown in FIG. 5, the center of gravity/crystal table 27 has a center of gravity item and a scintillator crystal item. The coordinates of the center of gravity defined in the flood map are registered in the center of gravity item. In the scintillator crystal item, the identifier of the scintillator crystal corresponding to the associated center of gravity coordinates is registered. For example, centroid coordinates "(x1, z1)" are associated with scintillator crystal "C1". That is, the interaction event for which "(x1, z1)" is calculated as the center of gravity coordinates is considered to have occurred in the scintillator crystal "C1". The detection processor 13 can identify the identifier of the scintillator crystal by comparing the barycenter coordinates calculated by the barycenter calculation function 23 with the barycenter/crystal table 27.

In the model update function 518, the processing circuit 51 updates the parameters of the machine learning model 531 based on the crystal divisions corrected by the crystal division correction function 516, and generates the updated machine learning model 531. The parameters are updated by reinforcement learning on the machine learning model 531.

In the display control function 519, the processing circuit 51 displays various information on the display 52. For example, the processing circuit 51 processes a PET image reconstructed by the reconstruction function 511, a flood map obtained by the flood map acquisition function 514, a crystal segment generated by the crystal segment generation function 515, and a crystal segment modified by the crystal segment correction function 516. Display the next crystal classification, etc.

The display 52 displays various information under the control of the processing circuit 51 in the display control function 519. As the display 52, for example, a CRT display, a liquid crystal display, an organic EL display, an LED display, a plasma display, or any other display known in the art can be used as appropriate.

The memory 53 is a storage device such as a ROM (Read Only Memory), a RAM (Random Access Memory), an HDD (Hard Disk Drive), an SSD (Solid State Drive), or an integrated circuit storage device that stores various information. Further, the memory 53 may be a drive device or the like that reads and writes various information to/from a portable storage medium such as a CD-ROM drive, a DVD drive, or a flash memory. For example, memory 53 stores machine learning model 531. The memory 53 also stores single event data, coincidence event data, PET images, display images, flood maps, crystal classifications, and the like.

The input interface 54 receives various input operations from the user, converts the received input operations into electrical signals, and outputs the electrical signals to the processing circuit 51. Specifically, the input interface 54 is connected to input devices such as a mouse, keyboard, trackball, switch, button, joystick, touch pad, and touch panel display. The input interface 54 outputs an electrical signal to the processing circuit 51 according to an input operation to the input device. Furthermore, the input device connected to the input interface 54 may be an input device provided in another computer connected via a network or the like.

Next, an example of the operation of the PET apparatus 1 according to this embodiment will be described.

FIG. 6 is a diagram showing a typical flow of a series of processes related to generation of the center of gravity/crystal table 27 by the processing circuit 51 according to the present embodiment. FIG. 7 is a diagram schematically showing a typical flow of the process in FIG. 6. The processes shown in FIGS. 6 and 7 are typically performed at the time of product shipment, periodic inspection, daily inspection, etc. of the PET apparatus 1.

As shown in FIGS. 6 and 7, first, the processing circuit 51 executes the flood map acquisition function 514 (step SA1). In step SA1, PET imaging is performed to generate the flood map 71, and gamma rays are detected by the gamma ray detector 17 to be processed. As described above, PET imaging for generating the flood map 71 is performed by injecting gamma rays from an arbitrary gamma ray emitter into the gamma ray detector 17 to be processed in a situation where no subject is present in the FOV. . The processing circuit 51 acquires a flood map 71 regarding the gamma ray detector 17 to be processed, based on the output signal of the gamma ray detector 17 to be processed.

When step SA1 is performed, the processing circuit 51 executes the crystal section generation function 515 (step SA2). In step SA2, the processing circuit 51 applies the machine learning model 531 to the flood map 71 acquired in step SA1 to generate crystal sections in the flood map 71. Note that the machine learning model 531 used in step SA2 may be the updated machine learning model 531 generated in step SA7 regarding the gamma ray detector 17 that was the previous processing target.

After step SA2 is performed, the processing circuit 51 waits for an instruction to modify the crystal classification (step SA3). In step SA3, the processing circuit 51 provides a GUI (Graphical User Interface) that accepts modification instructions.

FIG. 8 is a diagram showing an example of a display screen (hereinafter referred to as a modification operation screen) including a GUI that accepts instructions for modifying crystal classification. The modification operation screen is a type of reception section that receives instructions for modifying crystal classification. The correction operation screen is displayed on the display 52 by the processing circuit 51. As shown in FIG. 8, a flood map FM3 on which the crystal section SL3 is superimposed is displayed on the correction operation screen. The crystal section SL3 is displayed via the input interface 54 so that its shape can be changed.

The user determines whether or not the crystal section SL3 is drawn correctly while observing the flood map FM3 and the crystal section SL3. On the correction operation screen, for example, a display field MS for a message prompting correction is displayed. As the message, for example, "Please correct the crystal classification if necessary" is displayed. If the user determines that the crystal section SL3 is not drawn correctly, the user issues a modification instruction to change the shape of the crystal section SL3 via the input interface 54.

When a correction instruction for crystal section SL3 is given (step SA3: YES), processing circuit 51 executes crystal section correction function 516 (step SA4). In step SA4, the processing circuit 51 modifies the crystal section SL3 according to the modification instruction. The modification is reflected in the modification operation screen IS1 in real time.

FIG. 9 is a diagram showing an example of the modification of crystal section SL4 performed in step SA4. As shown in FIG. 9, crystal section SL4 is superimposed on flood map FM4 on the correction operation screen. In the local region R1 of the flood map FM4, the high frequency region of the center of gravity is spread over a relatively wide range, and it can be said that the accuracy of the crystal classification SL4 is relatively poor. In this case, the user issues an instruction via the input interface 54 to shift the crystal section SL4 within the local region R1 to the left. Correction instructions are issued in this manner.

As shown in FIG. 8, the correction operation screen further selectably displays a return button B1, a confirmation button B2, and a strengthen button B3. When the return button B1 is pressed, the processing circuit 51 returns the crystal section SL3 modified by the user via the input interface 54 to the initial crystal section SL3.

When a correction instruction is given and the confirm button B2 is pressed, the processing circuit 51 selects the crystal segment SL3 displayed on the correction operation screen at the time when the confirm button B2 is pressed, as the crystal to be used in the center of gravity/crystal table 27. Confirmed as a classification. The processing circuit 51 then executes the table generation function 517 (step SA5). In step SA5, the processing circuit 51 generates the center of gravity/crystal table 27 based on the flood map and the corrected crystal section. For example, as shown in FIG. 8, crystal section SL3 includes a plurality of closed regions. The plurality of closed regions correspond to the plurality of scintillator crystals actually mounted on the gamma ray detector 17, respectively. The plurality of closed regions are respectively associated with identifiers of the plurality of scintillator crystals. The processing circuit 51 identifies the closed region to which each pixel of the flood map belongs, and identifies the identifier of the scintillator crystal associated with the closed region. The processing circuit 51 registers the scintillator crystal identifier for each pixel of the flood map in the center of gravity/crystal table 27. As a result, the center of gravity/crystal table 27 is generated.

When step SA5 is performed, the processing circuit 51 determines whether or not to strengthen the machine learning model 531 used in step SA2 (step SA6). For example, when the confirm button B2 shown in FIG. 8 is pressed, the processing circuit 51 waits for the reinforcement button B3 to be pressed. If the reinforcement button B3 is pressed (step SA6: YES), the processing circuit 51 executes the model update function 518 (step SA7). In step SA7, the processing circuit 51 updates the parameters of the machine learning model 531 based on the crystal classification before modification and the crystal classification after modification. The parameters are updated by reinforcement learning on the machine learning model 531. The machine learning model 531 can learn manual corrections made by the user through reinforcement learning based on the crystal classification before correction and the crystal classification after correction. In this way, the accuracy of generating crystal sections based on the flood map can be improved.

On the other hand, if a correction instruction is not given in step SA3 (step SA3: NO), the processing circuit 51 uses the center of gravity/crystal table 27 based on the flood map acquired in step SA1 and the crystal classification generated in step SA2. generate. The method of generating the center of gravity/crystal table 27 is the same as step SA5.

If step SA7 is performed, if reinforcement learning is not performed (step SA6: NO), or if step SA8 is performed, the process in FIG. 6 ends.

The processing in FIGS. 6 and 7 is repeated for each gamma ray detector 17. The center of gravity/crystal table 27 for each gamma ray detector 17 is stored in the memory 21 of the detection processor 13 connected to the gamma ray detector 17. The detection processor 13 is enabled to use the centroid/crystal table 27 in PET imaging to identify the scintillator crystal in which the interaction event occurred.

Furthermore, by performing reinforcement learning based on corrections to the crystal classification for each correction, it becomes possible to cause the machine learning model 531 to generate crystal classifications that take into account manual corrections. That is, the machine learning model 531 according to this embodiment can recursively acquire generalization ability regarding manual correction. With this method, it is possible to improve the accuracy of the flood map classification work over time and autonomously. In dividing the flood map related to the gamma ray detector 17 to be processed next, the processing circuit 51 divides the crystals related to the flood map from the flood map related to the gamma ray detector 17 to be processed next according to the machine learning model 531 after the parameter update. Generate partitions. By dividing the flood map using such a machine learning model 531, the frequency of manual correction can be reduced, so that division can be performed simply and with high precision. Furthermore, by installing the machine learning model 531 that has been sufficiently learned into the PET apparatus 1, it becomes possible to adjust the gamma ray detector 17 on a daily basis, for example, without waiting for periodic inspection.

(Modified example)
In the above embodiment, the machine learning model 531 is a machine learning model whose parameters are learned so as to input a flood map and output a crystal classification. However, this embodiment is not limited to this. It is assumed that the machine learning model 531-2 according to the modification is a machine learning model in which parameters are learned so as to input a crystal section and output a crystal section with higher accuracy than the input crystal section. Hereinafter, a PET apparatus according to a modified example will be described. In the following description, components having substantially the same functions as those in this embodiment are denoted by the same reference numerals, and will be described repeatedly only when necessary.

FIG. 10 is a diagram showing a typical flow of a series of processes related to generation of a center of gravity/crystal table by the processing circuit 51 according to a modification of the present embodiment. FIG. 11 is a diagram schematically showing a typical flow of the process in FIG. 10. The processes in FIGS. 10 and 11, like the processes in FIGS. 6 and 7, are typically performed at the time of product shipment, periodic inspection, daily inspection, etc. of the PET apparatus 1.

As shown in FIGS. 10 and 11, first, the processing circuit 51 executes the flood map acquisition function 514 (step SB1). Step SB1 is similar to step SA1.

When step SB1 is performed, the processing circuit 51 executes the first crystal section generation function 515 (step SB2). In step SB2, the processing circuit 51 performs automatic segmentation processing on the flood map 81 acquired in step SB1 to generate crystal segments (hereinafter referred to as temporary crystal segments) 83 regarding the flood map. The automatic segmentation process uses a strict mathematical algorithm, a heuristic-defined algorithm, or a mixture thereof, which is different from the crystal segment generation process using the machine learning model 531-2. The automatic segmentation process is defined by a fixed procedure for generating pseudocrystalline segments 83 from the flood map, but does not incorporate manual modifications.

After step SB2 is performed, the processing circuit 51 executes the second crystal section generation function 515 (step SB3). In step SB3, the processing circuit 51 applies the machine learning model 531-2 according to the modification to the flood map 81 acquired in step SB1 and the pseudo-crystal section 83 generated in step SB2 to A crystal section 85 is produced. The machine learning model 531-2 is a machine learning model in which parameters are learned so as to receive a flood map and a crystal classification as input and output a crystal classification with higher accuracy than the input crystal classification. A highly accurate crystal segment means that the shape is more similar to a crystal segment after manual correction by a user or the like, compared to a crystal segment generated by automatic segmentation processing. In the machine learning model 531-2, parameters are learned using a combination of a flood map, a crystal segment generated by automatic segmentation processing, and a manually corrected crystal segment as a learning sample. Parameters are updated so that the output crystal section when the flood map and the crystal section generated by the automatic sectioning process are input approximates the manually corrected crystal section.

After step SB3 is performed, the processing circuit 51 waits for an instruction to modify the crystal classification (step SB4). In step SB4, the processing circuit 51 provides a GUI that accepts modification instructions, as shown in FIG.

If an instruction to modify the crystal division is given (step SB4: YES), the processing circuit 51 executes the crystal division correction function 516 (step SB5). In step SB5, the processing circuit 51 modifies the crystal division according to the modification instruction.

When a correction instruction is given, the processing circuit 51 determines the corrected crystal section as the crystal section to be used in the center of gravity/crystal table 27. The processing circuit 51 then executes the table generation function 517 (step SB6). In step SB6, the processing circuit 51 generates a center of gravity/crystal table based on the flood map and the corrected crystal section. Step SB6 is similar to step SA5.

When step SB6 is performed, the processing circuit 51 determines whether or not to strengthen the machine learning model 531-2 used in step SB3 (step SB7). When strengthening (step SB7: YES), the processing circuit 51 executes the model update function 518 (step SB8). In step SB8, the processing circuit 51 updates the parameters of the machine learning model 531-2 based on the crystal classification before modification and the crystal classification after modification. The parameter update is performed by reinforcement learning on the machine learning model 531-2. The machine learning model 531 can learn manual corrections made by the user through reinforcement learning based on the crystal classification before correction and the crystal classification after correction. This makes it possible to improve the accuracy of crystal section generation based on the flood map.

On the other hand, if no correction instruction is given in step SB4 (step SB4: NO), the processing circuit 51 uses the center of gravity/crystal table 27 based on the flood map acquired in step SB1 and the crystal classification generated in step SB3. generate. The method of generating the center of gravity/crystal table 27 is the same as step SB6.

If step SB8 is performed, if reinforcement learning is not performed (step SB7: NO), or if step SB9 is performed, the processing in FIGS. 10 and 11 ends.

The processing in FIGS. 10 and 11 is repeated for each gamma ray detector 17. The center of gravity/crystal table 27 for each gamma ray detector 17 is stored in the memory 21 of the detection processor 13 connected to the gamma ray detector 17. The detection processor 13 is enabled to use the centroid/crystal table 27 in PET imaging to identify the scintillator crystal where the interaction event occurred. Note that the machine learning model 531-2 may be trained to input the pseudo-crystal classification and output the crystal classification. In this case, in step SB3, the processing circuit 51 applies the temporary crystal section 83 generated in step SB2 to the machine learning model 531-2 to generate the crystal section 85.

(others)
In at least one embodiment described above, the medical information processing device is a computer built into a PET device. However, this embodiment is not limited to this. For example, a computer separate from the PET apparatus may be used. Such a computer includes, for example, a computer used in the manufacturing process of the PET apparatus 1 to generate a center of gravity/crystal table. In the following description, components having substantially the same functions as those in the above embodiment are denoted by the same reference numerals, and will be described repeatedly only when necessary.

FIG. 12 is a diagram showing an example of the configuration of a medical information processing apparatus 50 according to another example of the present embodiment. As shown in FIG. 12, the medical information processing device 50 includes a processing circuit 51, a display 52, a memory 53, an input interface 54, and a communication interface 55. For example, data communication between processing circuit 51, display 52, memory 53, input interface 54, and communication interface 55 is performed via a bus.

The processing circuit 51 has a processor such as a CPU or GPU as a hardware resource. The processing circuit 51 implements a flood map acquisition function 514, a crystal section generation function 515, a crystal section correction function 516, a table generation function 517, a model update function 518, and a display control function 519. The communication interface 55 is an interface for data communication with the PET apparatus for which the center of gravity/crystal table is to be generated. In the flood map acquisition function 514, the processing circuit 51 acquires a flood map from the PET apparatus for which the center of gravity/crystal table is to be generated, via the communication interface 55 or the like.

According to the above configuration, it is possible to realize at least the crystal segment generation function 515, the crystal segment correction function 516, the table generation function 517, and the model update function 518 in the medical information processing apparatus 50, which is a computer independent from the PET apparatus. become. This makes it possible for one medical information processing device 50 to perform tasks such as dividing a flood map for a plurality of PET devices. At this time, the model update function 518 performs reinforcement learning on the machine learning model 531 (or the machine learning model 531-2 according to the modified example) based on the correction of the crystal classification by the user. It becomes possible to efficiently generate a center of gravity/crystal table while strengthening the machine learning model 531-2).

(Application example 1)
In the above embodiment, the machine learning model 531 typically inputs and outputs the flood map in existing units such as detector packs or detector modules. However, this embodiment is not limited to this. The machine learning model according to Application Example 1 is trained on one flood map in small area (patch) units. The PET apparatus according to Application Example 1 will be described below. In the following description, components having substantially the same functions as those in this embodiment are denoted by the same reference numerals, and will be described repeatedly only when necessary.

The machine learning model according to Application Example 1 consists of a flood map in which a flood map is divided into small regions (hereinafter referred to as a divided flood map) and a division in which a section corresponding to the flood map is divided into small regions (hereinafter referred to as a divided flood map). ). By realizing the crystal division generation function 515, the processing circuit 51 divides the flood map to be processed, generates a plurality of divided flood maps to be processed, and applies the machine learning model to the plurality of divided flood maps to be processed. A plurality of divided sections are generated, each corresponding to a plurality of divided flood maps to be processed. The processing circuit 51 then integrates the plurality of divided sections to be processed and generates a crystal section to be processed.

FIG. 13 is a diagram schematically showing a typical flow of the process of generating crystal sections 73 from flood map 71 by processing circuit 51 according to application example 1. As shown in FIG. 13, the processing circuit 51 acquires a flood map 71 to be processed regarding an existing unit such as a detector pack or a detector module by implementing a flood map acquisition function 514. Hereinafter, the size of the existing unit flood map, etc. will be referred to as the existing size. Next, the processing circuit 51 executes division processing by realizing the crystal section generation function 515 (step SC1). In step SC1, the processing circuit 51 divides the flood map 71 to be processed and generates a plurality of divided flood maps 711 having mutually different positions.

FIG. 14 is a diagram showing the relationship between the flood map 71 and the divided flood map 711. As shown in FIG. 14, the size of the divided flood map 711 may be any size as long as it is smaller than the size of the original flood map 71. For example, if the flood map 71 has the size of 16 scintillator crystals in the z-axis direction and 8 scintillator crystals in the x-axis direction, the divided flood map 711 has the size of 16 scintillator crystals in the z-axis direction and 8 scintillator crystals in the x-axis direction. Set to size. The sizes of the plurality of divided flood maps 711 divided from one flood map 71 are all set to the same size. Adjacent divided flood maps 711 are divided so as to overlap in a partial area 712. The number of divided flood maps that share a partial area 712 may be two or three or more. By dividing the plurality of divided flood maps 711 so that two or more divided flood maps 711 overlap, it is possible to avoid partial omission of crystal sections generated in subsequent processing.

For example, the processing circuit 51 generates the divided flood map 711 by applying a patch having the same size as the divided flood map 711 to the generation target area of the divided flood map 711 of the flood map 71 and cutting it out. The processing circuit 51 generates a divided flood map 711 for each patch application position while shifting the patch application position. The patch application position may be automatically determined or may be set by the user via the input interface 54 or the like. Hereinafter, the size of the divided flood map 711 etc., ie, the size of the patch, will be referred to as patch size.

When the divided flood maps 711 are generated, the processing circuit 51 applies the machine learning model 531-3 to each divided flood map 711 of the patch size to generate divided crystal sections 731 of the patch size. In the processing circuit 51, the parameters of the machine learning model 531-3 are learned in advance so as to input the divided flood map 711 and output the divided crystal section 731. Typically, the same model is used as the machine learning model 531-3 regardless of the position of the divided flood map 711 in the flood map 71. In this way, the machine learning model 531-3 can be generalized by learning on a patch-by-patch basis. Note that the machine learning model 531-3 may be prepared depending on the position of the divided flood map 711 in the flood map 71.

Next, the processing circuit 51 integrates the plurality of split crystal sections 731 of the patch size to generate a crystal section 73 of the existing size (step SC2). In step SC2, the processing circuit 51 generates a crystal section 73 of an existing size by pasting together a plurality of patch-sized crystal sections 731 according to the position of each divided crystal section.

FIG. 15 is a diagram schematically showing the integration process performed in step SC2. As shown in FIG. 15, for example, two split crystal sections 731 are generated. The two divided crystal sections 731 shown in FIG. 15 correspond to the two divided flood maps 711 shown in FIG. 14, respectively. The processing circuit 51 bonds the two divided crystal sections 731 together according to the position of each divided crystal section 731. Since the two divided flood maps 711 shown in FIG. 14 overlap in a partial area 712, the two divided crystal sections 731 are pasted together so as to overlap in a partial area 732 corresponding to this partial area 712. . In addition, in FIG. 15, for visual distinction, the dividing line 7311 of the upper divided crystal section 731 is represented by a broken line, and the dividing line 7312 of the lower divided crystal section 732 is represented by a solid line. In addition, the partial area 732 will also be referred to as an overlapping area 732 hereinafter.

In some areas 732, the dividing lines drawn in each divided crystal section 731 may not match. For example, as shown in FIG. 15, it is ideal that the dividing line 7311 of the upper divided crystal section 731 and the dividing line 7312 of the lower divided crystal section 732 coincide in the overlapping region 732, but they may deviate. be. The processing circuit 51 performs averaging processing between the upper crystal division 731 and the lower crystal division 732 in the overlap region 732 in order to correct the mismatch between the dividing lines 7311 and 7312 in the overlap region 732. In the averaging process, a line is drawn at an average position between the division line 7311 and the division line 7312, in other words, at a substantially midway position. Through the averaging process, continuity of the division lines in each crystal division 731 can be ensured. As described above, by performing the bonding process and the averaging process on all the divided crystal sections 731, crystal sections 73 of the existing size are generated.

Once crystal section 73 is generated, steps SA3-SA7 shown in FIG. 6 are performed.

As described above, according to Application Example 1, learning of the machine learning model 531-3 can be generated in patch units smaller than the entire flood map. By learning on a patch-by-patch basis, the size of the input/output data of the machine learning model 531-3 can be reduced, thereby increasing the drawing accuracy of crystal sections.

(Application example 2)
In the above embodiment, the machine learning model 531-2 typically inputs and outputs crystal sections in existing units such as detector packs or detector modules. However, this embodiment is not limited to this. The machine learning model according to Application Example 2 learns one crystal section in units of small regions (patch). The PET apparatus according to Application Example 2 will be described below. In the following description, components having substantially the same functions as those in this embodiment are denoted by the same reference numerals, and will be described repeatedly only when necessary.

The machine learning model according to Application Example 2 is trained based on a patch size crystal classification and a patch size crystal classification having higher accuracy than the crystal classification. By implementing the crystal segment generation function 515, the processing circuit 51 applies automatic segmentation processing to the flood map to be processed, generates pseudo-crystal segments for the flood map to be processed, and performs patching from the generated pseudo-crystal segments. Generate crystal divisions of size. Next, the processing circuit 51 generates a crystal segment with a high precision patch size based on the machine learning model and the temporary crystal segment of the patch size by realizing the crystal segment generation function 515, and generates a crystal segment with a high precision patch size. A crystal section of the existing size is generated by integrating the crystal section of the size into a temporary crystal section. A highly accurate crystal segment means that the shape is more similar to a crystal segment after manual correction by a user or the like, compared to a crystal segment generated by automatic segmentation processing.

FIG. 16 is a diagram schematically showing a typical flow of the process of generating crystal sections 85 from flood map 81 by processing circuit 51 according to application example 2. As shown in FIG. 16, first, the processing circuit 51 acquires the flood map 81 to be processed by implementing the flood map acquisition function 514. Next, the processing circuit 51 performs automatic segmentation processing on the acquired flood map 81 by implementing the first crystal segment generation function 515, and generates a temporary crystal segment 83 of an existing size regarding the flood map ( Step SD1). Step SD1 is the same as step SB2 in FIGS. 10 and 11. Next, the processing circuit 51 cuts out the temporary crystal section 83 as a patch, and generates a cut-out crystal section 831 of the patch size (step SD2). Similar to Application Example 1, the patch has a smaller size than the size of the detector pack or detector module unit. The patch is applied to the area of the temporary crystal section 83 that requires modification. The patch application position is specified by the user via the input interface 54, for example. For example, the processing circuit 51 provides a GUI that accepts an instruction to designate an application position.

FIG. 17 is a diagram illustrating an example of a display screen (referred to as a cutout operation screen) including a GUI that accepts an instruction to designate a patch application position. As shown in FIG. 17, the cutting operation screen is displayed on the display 52 by the processing circuit 51. As shown in FIG. 17, a flood map FM3 on which the temporary crystal section SL3 is superimposed is displayed on the cutout operation screen. A patch P1 that can be moved via the input interface 54 is displayed on the cutout operation screen. On the clipping operation screen, for example, a display field MS2 for a message prompting clipping using patch P2 is displayed. As the message, for example, "Please cut out the modified part with a patch" is displayed. The user observes the temporary crystal section SL3, finds a portion where the accuracy of the temporary crystal section SL3 is desired to be improved, and moves the patch P1 to the relevant portion via the input interface 54. After moving the patch P1, the user presses the cutout button B5.

When the cutout button B5 is pressed, the processing circuit 51 cuts out the region surrounded by the patch P1 of the temporary crystal section SL3, and generates a patch-sized cutout crystal section 831. Next, the processing circuit 51 performs a correction process on the cut-out crystal section 831 using the machine learning model 531-4. The machine learning model 531-4 is trained in advance based on input data that is a crystal classification of a patch size and teacher data that is a crystal classification of a patch size that is more accurate than the crystal classification. The processing circuit 51 applies the patch-sized cut-out crystal section 831 to the machine learning model 531-4 to generate a more accurate patch-sized cut-out crystal section 832.

As shown in FIG. 17, the cutout operation screen further displays a return button B4 and an add button B6 in a selectable manner. When the add button B6 is pressed, the processing circuit 51 receives an instruction to designate a further application position of the patch P2. When the designation instruction is given, the processing circuit 51 applies the patch P2 to a further application position, generates a cutout crystal section 831, and applies the generated cutout crystal section 831 to the machine learning model 531-4 to determine the patch size. A cut-out crystal section 832 is generated. When the user determines that he/she wants to cancel the modification made by the machine learning model 531-4, the user presses the return button B4 via the input interface 54. When the return button B4 is pressed, the processing circuit 51 displays the temporary crystal section SL3 before application of the machine learning model 531-4.

After generating the cut-out crystal section 832, the processing circuit 51 integrates the patch-sized cut-out crystal section 832 into the existing-sized temporary crystal section 83 to generate a high-precision crystal section 85 of the existing size (step SD3).

FIG. 18 is a diagram showing the relationship between the cut-out crystal section 832 of the patch size and the temporary crystal section 83 of the existing size, which are integrated in the integration process of FIG. 16. As shown in FIG. 18, the processing circuit 51 attaches the cut-out crystal section 832 to the temporary crystal section 83 of the existing size according to the position of the cut-out crystal section 832, and performs averaging processing on the cut-out crystal section 832 and the temporary crystal section 83. conduct. The bonding process and the averaging process are the same as in Application Example 1. This produces a highly accurate crystal section 85.

Once crystal section 85 is generated, steps SB4-SB8 shown in FIG. 10 are performed.

As described above, according to Application Example 2, the machine learning model 531-4 can be trained in patch units that are smaller than the entire flood map. By learning on a patch-by-patch basis, the size of input and output data of the machine learning model 531-4 can be reduced, thereby improving the drawing accuracy of crystal sections.

Further, in the above embodiment, the machine learning model 531 (or the machine learning models 531-2, 531-3, 531-4 according to modified examples) is a deep neural network or a parameterized composite function stored in the memory 53. The processing circuit 51 generates crystal sections according to the machine learning model 531 (or the machine learning models 531-2, 531-3, and 531-4 according to modified examples). However, this embodiment is not limited to this. The processing circuit 51 may generate crystal segments in the flood map to be processed from the flood map to be processed by using a machine learning structure constructed to input a flood map and output crystal segments. , may be in any form. The machine learning structure according to this embodiment includes a software configuration of a machine learning model that describes a function to be realized by a processor, and a hardware configuration of a circuit that implements the function described by the machine learning model.

For example, the processing circuit 51 may be a programmable logic device (Application Specific Integrated Circuit (ASIC) having a machine learning structure constructed to input a flood map and output a crystal section), a programmable logic device (for example, a simple It may also be equipped with integrated circuits such as a simple programmable logic device (SPLD), a complex programmable logic device (CPLD), and a field programmable gate array (FPGA). The integrated circuit has a circuit configuration that realizes the same function as the machine learning model 531 whose parameters are learned so as to input a flood map and output a crystal section.

Similarly, the processing circuit 51 inputs a flood map and a crystal classification, and utilizes a machine learning structure constructed to output a crystal classification with higher accuracy than the input crystal classification. Any form may be used as long as it generates highly accurate crystal sections from the flood map and the crystal sections. For example, the processing circuit 51 is equipped with an integrated circuit such as an ASIC, a SPLD, a CPLD, and an FPGA having a machine learning structure constructed to input a flood map and a crystal classification and output a highly accurate crystal classification. Also good.

According to at least one embodiment described above, a flood map can be divided easily and accurately.

The word "processor" used in the above description means, for example, a CPU, a GPU, or an integrated circuit such as an ASIC, a SPLD, a CPLD, and an FPGA. A processor realizes its functions by reading and executing a program stored in a memory circuit. Note that instead of storing the program in the memory circuit, the program may be directly incorporated into the circuit of the processor. In this case, the processor realizes its functions by reading and executing a program built into the circuit. Furthermore, instead of executing a program, functions corresponding to the program may be realized by combining logic circuits. Note that each processor of this embodiment is not limited to being configured as a single circuit for each processor, but may also be configured as a single processor by combining multiple independent circuits to realize its functions. good. Furthermore, a plurality of components in FIGS. 1 and 12 may be integrated into one processor to realize its functions.

Although several embodiments of the invention have been described, these embodiments are presented by way of example and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, substitutions, and changes can be made without departing from the gist of the invention. These embodiments and their modifications are included within the scope and gist of the invention as well as within the scope of the invention described in the claims and its equivalents.

1 PET device 10 Gantry 11 Detector ring 13 Detection processor 15 Coincidence circuit 17 Gamma ray detector 19 Processing circuit 21 Memory 23 Center of gravity calculation function 25 Energy calculation function 27 Center of gravity/crystal table 30 Bed 50 Medical information processing device (console)
51 Processing circuit 52 Display 53 Memory 54 Input interface 55 Communication interface 511 Reconfiguration function 512 Image processing function 513 Imaging control function 514 Flood map acquisition function 515 Crystal division generation function 516 Crystal division correction function 517 Table generation function 518 Model update function 519 Display Control function 531 Machine learning model

Claims (14)

an acquisition unit that acquires a flood map to be processed that indicates the spatial distribution of the frequency of fluorescence derived from gamma rays;
a machine learning model that outputs divisions in a flood map for identifying scintillator crystals in which the fluorescence has occurred based on a flood map showing a spatial distribution of frequency of fluorescence derived from gamma rays; a division generation unit that generates divisions in the flood map to be processed;
a correction unit that corrects the classification output from the machine learning model;
an updating unit that updates parameters of the machine learning model based on the modified classification and generates an updated machine learning model,
The division is defined by the coordinate system of the flood map and is line data indicating a division of the scintillator crystal.
Medical information processing equipment. The medical information according to claim 1, wherein the classification generation unit outputs a classification in the processing target flood map or other flood map from the processing target flood map or other flood map according to the updated machine learning model. Processing equipment. The medical information processing apparatus according to claim 1, further comprising a reception unit that receives an instruction to modify the classification output from the machine learning model. The medical information processing apparatus according to claim 1, further comprising a display unit that displays a GUI on a display for instructing the updating unit to update parameters based on the modified classification. The medical information processing according to claim 1, further comprising an association data generation unit that generates association data between a position in the flood map and an identifier of a scintillator crystal corresponding to the position, based on the classification and the flood map. Device. The medical information processing device according to claim 1, further comprising a storage unit that stores the machine learning model. The machine learning model is learned based on a plurality of divided flood maps in which the flood map is divided into a plurality of regions, and a plurality of divided sections in which a section corresponding to the flood map is divided into a plurality of regions. The medical information processing device according to claim 1. The division generation unit divides the flood map to be processed to generate a plurality of divided flood maps to be processed, and applies the machine learning model to the plurality of divided flood maps to be processed to generate the plurality of divided flood maps to be processed. The medical information processing apparatus according to claim 7, wherein the medical information processing apparatus generates a plurality of divided sections respectively corresponding to a plurality of divided flood maps. 9. The medical information processing apparatus according to claim 8, wherein the division generation unit generates the division of the processing target by integrating a plurality of divided divisions of the processing target. The medical information processing apparatus according to claim 7, wherein at least some of the plurality of divided flood maps are set to overlap with each other. The division generation unit divides the flood map to be processed so that at least a portion of the area overlaps to generate a plurality of divided flood maps to be processed, and divides the flood map to be processed by the machine into the plurality of divided flood maps to be processed. The medical information processing apparatus according to claim 10, wherein a learning model is applied to generate a plurality of divided sections respectively corresponding to the plurality of divided flood maps to be processed. The medical device according to claim 11, wherein the division generation unit generates the division of the processing target by integrating the plurality of divisions of the processing target, and averages two or more divisions that share the partial area. Information processing device. an acquisition unit that acquires a flood map to be processed that indicates the spatial distribution of the frequency of fluorescence derived from gamma rays;
a first division generation unit that applies a predetermined algorithm to the flood map to be processed to generate a first division regarding the flood map to be processed for identifying the scintillator crystal in which the fluorescence has been generated;
A machine learning model that outputs a second classification that is more accurate than the first classification based on the first classification, and a machine learning model that outputs a second classification that is more accurate than the first classification; a second division generation unit that generates a second division regarding the flood map to be processed;
A medical information processing device comprising: Utilizing a machine learning structure constructed to input a flood map indicating the spatial distribution of the frequency of fluorescence derived from gamma rays and output a division in the flood map for identifying the scintillator crystal in which the fluorescence has occurred. , generating a division in the flood map to be processed from the flood map to be processed;
modifying the classification output from the machine learning structure ;
updating parameters of the machine learning structure based on the modified classification to generate an updated machine learning structure ;
Here, the division is data of lines indicating divisions of the scintillator crystal defined by the coordinate system of the flood map.
Medical information processing equipment.