Nuclear medicine diagnosis device, image processing method and program
Patent Information
- Application Number
- JP2022205153
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2022-02-28
- Filing Date
- 2022-12-22
- Publication Date
- 2025-12-25
AI Technical Summary
Existing PET image reconstruction methods face challenges in accurately and efficiently correcting for scattering effects, with Monte Carlo simulations being too slow and model-based methods being complex and costly, while deep learning techniques struggle to directly estimate scatter from emission or attenuation sinograms.
A nuclear medicine diagnostic apparatus using a trained deep convolutional neural network (DCNN) processes attenuation-corrected data and nuclear medicine images to directly estimate scatter sinograms and generate scatter-corrected images, employing a hybrid approach that combines image-based and sinogram-based corrections.
The hybrid deep learning method significantly improves image quality by reducing scatter effects, achieving faster computation times and more accurate scatter correction compared to traditional methods, with the liver region showing enhanced uniformity and overall image clarity.
Smart Images
Figure 00000000_0000_ABST
Abstract
Description
[Technical Field]
[0001] Embodiments disclosed herein and in the drawings relate to nuclear medicine diagnostic devices, image processing methods, and programs. [Background technology]
[0002] Scattering is one of the main factors that degrade the quality of reconstructed PET (Positron Emission Tomography) images. Currently, there are two main methods for scattering correction: Monte Carlo simulation and model-based scattering simulation.
[0003] Monte Carlo simulations are highly accurate and include all scattering events, but they are too slow for use in many real-time environments. In contrast, known model-based single-scatter simulations are relatively fast, but they still represent a time-consuming part of the reconstruction process. Furthermore, they only predict the sinogram for a single scattering event, requiring additional prediction steps for multiple scattering events. Recently, it has been demonstrated that secondary scattering events can also be modeled as part of model-based simulations, and that the resulting sinograms are very close to the complete scattering sinogram. However, this method suffers from the problems of being very complex and computationally expensive.
[0004] Deep Convolutional Neural Networks (DCNNs) are widely used as machine learning techniques for image processing. Generally, DCNNs involve multiple steps such as convolution, stride, pooling, and padding. During convolution, the network uses a kernel to extract the desired features from the input data.
[0005] Two techniques using DCNNs for PET scattering estimation are known: (1) a technique for estimating multiple scattering sinograms from a single scattering sinogram, and (2) a technique for directly estimating scattering sinograms from emission attenuation sinograms. The latter method is difficult because scattering is not physically generated by emission or attenuation sinograms, and there is no direct relationship between these sinograms and scattering sinograms. Furthermore, intrinsic image features that affect scattering cannot be detected from these sinograms. [Prior art documents] [Patent Documents]
[0006] [Patent Document 1] U.S. Patent No. 11145055 [Overview of the project] [Problems that the invention aims to solve]
[0007] One of the problems that the embodiments disclosed in this specification and drawings aim to solve is to improve image quality. However, the problems that the embodiments disclosed in this specification and drawings aim to solve are not limited to the above problem. Problems corresponding to the effects of each configuration shown in the embodiments described later can also be positioned as other problems. [Means for solving the problem]
[0008] The nuclear medicine diagnostic apparatus according to the embodiment includes a processing circuit. The processing circuit receives attenuation correction data of a subject and a first nuclear medicine image of the subject, which is a nuclear medicine image including scattering effects, and further inputs the attenuation correction data and the first nuclear medicine image into a trained machine learning-based system to output a first scattering-based output that includes data corresponding to the scattering effects in the first nuclear medicine image. The trained machine learning-based system is trained by inputting (1) attenuation correction training data and (1b) a training nuclear medicine image including scattering effects as input datasets, and (2) corresponding sets of scattering-based training outputs in which the scattering effect is reduced compared to the training nuclear medicine image, wherein the first scattering-based output includes at least one of (a) a scattering sinogram showing estimated scattering in the first nuclear medicine image, (b) a set of angle projections of scattering, and (c) a scattering-corrected image in which the scattering is reduced compared to the first nuclear medicine image. [Brief explanation of the drawing]
[0009] [Figure 1A] Figure 1A shows a machine learning-based system (e.g., DCNN) that is trained to directly generate estimated scattering sinograms from corresponding pairs of nuclear medicine training image data (e.g., images from multiple PET scans) and training attenuation correction data (e.g., attenuation images / maps and / or data from a 3D model of radiation attenuation) using estimated scattering data generated by at least one other method (e.g., Monte Carlo simulation and / or a model-based method). [Figure 1B] Figure 1B shows a trained machine learning-based system (e.g., DCNN) that directly generates estimated scattering sinograms from nuclear medicine images (e.g., images from PET scans) and (multiple) types of attenuation correction data (e.g., data from attenuated images / maps and / or 3D models of radiation attenuation) used to train the trained machine learning-based system. [Figure 1C]FIG. 1C is a diagram showing an exemplary image for an estimated scatter sinogram, a nuclear medicine image, and the method and system of FIG. 1B utilizing attenuation correction data. [Figure 1D] FIG. 1D is a diagram showing a set of trained DCNNs that directly generate the scatter estimation angle-by-angle projections from a nuclear medicine image (e.g., an image from a PET scan) and the type of attenuation correction data (e.g., data from an attenuation image / map and / or a three-dimensional model of radiation attenuation) used in the training of a set of trained DCNNs. [Figure 2] FIG. 2 is a diagram showing a data flow in which a trained DCCN creates a set of scatter correction images from (a) a nuclear medicine image and (b) attenuation correction data, then creates a scatter sinogram by forward projection, and applies forward projection to a set of difference images generated from the nuclear medicine image and the set of scatter correction images. [Figure 3] FIG. 3 is a diagram showing the process in which a difference image is generated from the nuclear medicine image of FIG. 2 and one of the set of scatter correction images. [Figure 4] FIG. 4 is a diagram showing the process of applying forward projection to the difference image of FIG. 3 to generate a pseudo scatter sinogram for scatter correction. [Figure 5] FIG. 5 is a diagram showing one exemplary architecture of a 3D deep neural network having a PET / CT multi-channel input according to an embodiment. [Figure 6] FIG. 6 is a diagram showing different images of cancer patients, including imaging of patients (1) without scatter correction (Non-SC), (2) with scatter correction using single scatter simulation (SSS-SC), and (3) with deep scatter correction (DPET-SC). [Figure 7A] FIG. 7A is a diagram showing a perspective view of a PET device according to an embodiment of the present disclosure. [Figure 7B] FIG. 7B is a diagram showing a schematic view of a PET device according to an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE INVENTION
[0010] Hereinafter, embodiments of a nuclear medicine diagnostic apparatus, an image processing method, and a program will be described in detail while referring to the drawings.
[0011] In nuclear medicine imaging, the image quality of a PET scan deteriorates due to the influence of scatter. To correct for scatter, a machine learning-based correction system is utilized, and in one embodiment, a real-time correction process using a DCNN is described herein. In one embodiment, a scatter sinogram is directly estimated from emission attenuation correction data using a DCNN. In another embodiment, after estimating a scatter-corrected image using a DCNN, the scatter sinogram is calculated by forward projection.
[0012] In any of the embodiments, the DCNN can be implemented as a structure such as a U-Net, but the network structure is not limited to this. Other different combinations of basic neural network elements such as convolution operations, activation functions, max pooling, batch normalization, etc. can also be used. In the training of the network, a loss function that reflects the similarity between the predicted image (or sinogram) and the ground truth scatter-corrected image (or scatter sinogram) is minimized. These network parameters form a trained network that is used with a new set of images. Various similarity metrics can be applied to measure the similarity between the prediction and the ground truth. This similarity metric includes, but is not limited to, root mean square error, weighted sum of intensity differences, cross-correlation, adversarial loss, and mutual information between image histograms.
[0013] Figure 1A shows a machine learning-based system (e.g., DCCN) that is trained to estimate scattering from a set of training image data representing multiple time-of-flight (TOF) reconstructions without scattering correction, and a corresponding set of training attenuation-corrected data (e.g., generated from a 3D model of radiation attenuation or a set of attenuation images). In one embodiment, the set of training attenuation-corrected data is obtained from a number of helical computed tomography (CT) scans that are input to the machine learning-based system along with PET image data. Alternatively, the attenuation-corrected data may be corresponding data obtained from a 3D model of radiation attenuation. Before DCCN training, the attenuation-corrected data must be registered with the nuclear medicine images. Next, the machine learning-based system is trained to emulate the scattering information of the target generated using an alternative scattering estimation method (e.g., Monte Carlo simulation or a model-based method).
[0014] Subsequently, the trained machine learning-based system (e.g., shown in Figures 1B and 1C) can generate an estimated scattering sinogram from the subsequent input image data and the corresponding attenuation-corrected data of the type used to train the trained system. An illustrative image is shown in relation to the data flow diagram in Figure 1C.
[0015] In an alternative embodiment, the single DCNN in Figure 1B is replaced with a series of angle-specific DCNNs to reduce the network dimensionality compared to calculating all imaging angles using a single network. Figure 1D shows a set of trained DCNNs that directly generate estimated angle-specific projections of scattering from nuclear medicine images (e.g., images from PET scans) and the type of attenuation correction data used to train the trained DCNN. The individual projections of scattering from various angles can be combined to generate a scattering sinogram. For non-time-of-flight (TOF) scattering, the projection is two-dimensional. For TOF scattering, the projection is three-dimensional.
[0016] In an alternative embodiment, instead of creating a machine learning-based system to generate scattered sinograms, the machine learning-based system is trained to generate scatter-corrected images. FIG. 2 shows a data flow diagram where a trained DCCN creates a set of scatter-corrected images from (a) a nuclear medicine image and (b) attenuation correction data. Next, forward projection is applied to a set of difference images generated from the nuclear medicine image and the set of scatter-corrected images to generate at least a set of scatter-corrected sinogram data Px s is generated. Next, at least a set of scatter-corrected sinogram data Px s can be used for any number of reconstructions, including Maximum Likelihood Expectation Maximization (MLEM) reconstruction. The reconstruction is not limited to this.
[0017] To train an untrained machine learning-based system to generate corresponding scatter-corrected images, the system seeks to determine a network that generates a scatter distribution that minimizes the cost function given by Equation (1) below.
[0018] [Number]
[0019] Here, f θ represents a network that maps an un-scatter-corrected image x nsc to a ground truth scatter-corrected image x sc through supervised training. Alternatively, other cost functions may be used.
[0020] Here, x μ is another input channel that contains attenuation information to assist in scatter estimation / modeling. The difference image between x nsc and f θ (x nsc ,x μ ) is defined as the scatter distribution x s .
[0021] Here, the following Equation (2) holds.
[0022]
number
[0023] Here, y (bar symbol) represents the measured count of the sinogram. P is the system matrix representing the forward projection, x sc This is a ground truth scattering corrected image, x s Px is the scattering distribution described above, and r (bar symbol) is the expected number of simultaneous occurrences due to random chance. s The difference between the model-based scattering sinogram and the model-based scattering sinogram is assumed to be located in the zero space of the back projector and does not affect the reconstruction. This allows for a scattering-corrected reconstruction as shown in equation (3) below.
[0024]
number
[0025] In the PET maximum likelihood EM algorithm, the iterative update formula is given by equation (4) below.
[0026]
number
[0027] Here y i is the measurement count in the i-th LOR, and x j k This is the estimated k-th activation in the j-th voxel.
[0028] Scattering correction in iterative reconstruction and estimated scattering s from the proposed method above i Taking this into consideration, the iterative update formula with scattering correction is as follows (5):
[0029]
number
[0030] Figure 3 is a data diagram showing a portion of Figure 2. Figure 3 illustrates a difference image generated from a nuclear medicine image and one of the images in the pair of scattering-corrected images from Figure 2. In the data diagram of Figure 4, the difference image from Figure 3 is projected forward to generate a pseudo-scattering sinogram, and scattering correction is performed.
[0031] To address timing and performance issues in known techniques, image-based scattering estimation uses a smaller data size (3D images vs. 4D / 5D sinograms) than sinogram-based scattering estimation. Using scattering sinograms provides more flexible scattering correction compared to directly outputting the scattered-corrected image.
[0032] In one embodiment, a machine learning-based system consists of a neural network, and a trained machine learning-based system consists of a trained neural network.
[0033] In another embodiment, the machine learning-based system consists of multiple neural networks, at least one of which utilizes angular scattering correction of forward projection, as shown in Figure 1D.
[0034] Figure 5 shows one exemplary architecture of a 3D deep neural network having PET / CT multichannel input according to one embodiment of the present disclosure. It shows multiple hidden layers and one output layer. As shown in the figure, a set of PET images and attenuation information are applied as input to the neural network. The neural network in the figure shows the image size used in each layer and the number of connections used for interconnecting each layer. The type of operation performed in each layer is also shown. As seen in the exemplary embodiment, the image size is reduced during processing of many layers and then re-enlarged in subsequent layers to improve the speed of learning and testing. However, it is not necessary to reduce the image size or reduce it by the amount that is displayed.
[0035] Figure 6 shows different images of cancer patients, including (1) images of patients without scattering correction (Non-SC), (2) images with scattering correction using single scattering simulation (SSS-SC), and (3) images with deep scattering correction (DPET-SC).
[0036] Compared to direct estimation of scattering sinograms using deep neural networks, image-based scattering estimation is more efficient because time-of-flight PET sinograms are 5-dimensional while images are 3-dimensional. Compared to pure image-based deep scattering correction, hybrid methods are expected to further reduce bias.
[0037] Deep scattering correction was validated using fluoride deoxyglucose (FDG) scans from a Canon® Cartisian® PET / CT scanner. A pre-trained deep scattering network was applied to uncorrected images to generate corrected images. Network prediction was performed using linear regression R. 2 This corresponds to a scattering-corrected image using a single scattering simulation with a value of 0.97.
[0038] As the images and associated measurements demonstrate, the deep scattering method of this disclosure yielded good image quality. In the deep scattering-corrected image, the liver region is more uniform compared to the single-scatter simulation-corrected image. Furthermore, the computation time for deep scattering estimation was only 2 seconds, which is significantly faster than the single-scatter simulation method.
[0039] It can be understood that the above technology can be incorporated into various systems. In one embodiment, the above technology can be incorporated into a PET system. Figures 7A and 7B show a non-limiting example of a PET apparatus 400 that can carry out the method described herein. The PET apparatus 400 comprises a number of gamma-ray detectors (GRDs) (e.g., GRD1, GRD2 to GRDN) each configured as a rectangular detector module.
[0040] Each GRD may include a two-dimensional array of individual detector crystals that absorb gamma rays and emit scintillation photons. These scintillation photons can be detected by a two-dimensional array of photomultiplier tubes (PMTs). The photomultiplier tubes are also placed within the GRD. An optical guide can be placed between the array of detector crystals and the PMTs.
[0041] Alternatively, scintillation photons can be detected by an array of silicon photomultipliers (SiPMs), and each detector crystal can have its own SiPM.
[0042] Each photodetector (e.g., PMT or SiPM) can generate an analog signal. This analog signal indicates when a scintillation event occurs and the energy of the gamma ray that generates the detection event. Furthermore, photons emitted from a single detector crystal can be detected by multiple photodetectors, and based on the analog signals generated by each photodetector, the detector crystal corresponding to the detection event can be determined, for example, using Anger logic and crystal decoding.
[0043] Figure 7B shows a schematic diagram of a PET scanner system having a gamma-ray photon counting detector (GRD) positioned to detect gamma rays emitted from the subject's obstetric blob (OBJ). The GRD can measure the timing, position, and energy corresponding to each gamma-ray detection. In one embodiment, the gamma-ray detectors are arranged in a ring, as shown in Figures 7A and 7B. The detector crystal can be a scintillator crystal in which individual scintillator elements are arranged in a two-dimensional array. The scintillator elements can be any known light-emitting material. The PTMs can be arranged so that light from each scintillator element is detected by multiple PTMs, enabling Anger calculation and crystal decoding of the scintillation event.
[0044] Figure 7B shows an example of the arrangement of the PET scanner 400, where the subject's obj (OBJ) to be imaged is placed on the table 416, and the GRD modules GRD1 to GRDN are arranged around the subject's OBJ and the table 416. The GRDs can be fixedly connected to a circular component 420 which is fixedly connected to the gantry 440. The gantry 440 houses many of the components of the PET imaging device. The gantry 440 of the PET imaging device also includes an opening through which the subject's OBJ and the table 416 can pass, and the GRDs can detect gamma rays emitted in the opposite direction from the subject's OBJ due to an extinction event, and the matching of gamma ray pairs can be determined using timing and energy information.
[0045] Figure 7B also shows the circuitry and hardware for acquiring, storing, processing, and distributing gamma-ray detection data. The circuitry and hardware include a processor 470, a network controller 474, a memory 478, and a Data Acquisition System (DAS) 476. The PET imaging system also includes data channels that transfer detection measurement results from the GRD to the DAS 476, processor 470, memory 478, and network controller 474. The DAS 476 can control the acquisition, digitization, and routing of detection data from the detector. In one embodiment, the DAS 476 controls the movement of the table 416. The processor 470 performs functions such as image reconstruction from the detection data, pre-processing of the detection data for reconstruction, and post-processing of the image data for reconstruction, as described herein.
[0046] In one embodiment, the processor 470 can be configured to perform various steps and variations thereof as described herein. The processor 470 may include a CPU that can be implemented as discrete logic gates, an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA), or other Complex Programmable Logic Device (CPLD). The FPGA or CPLD implementation may be coded in VHDL, Verilog, or any other hardware description language, and the code may be stored in electronic memory provided inside the FPGA or CPLD or as separate electronic memory. Furthermore, the memory may be non-volatile memory such as ROM, EPROM®, EEPROM®, or FLASH® memory. The memory may also be volatile memory such as static or dynamic RAM, and a processor such as a microcontroller or microprocessor may be provided to manage the electronic memory as well as the interaction between the FPGA or CPLD and the memory.
[0047] Alternatively, the CPU in processor 470 can execute a computer program comprising a set of computer-readable instructions that perform the various steps described herein, the program stored in any of the non-temporary computer-readable media, electronic memory and / or hard disk drives, CDs, DVDs, FLASH® drives or other known storage media. Furthermore, the computer-readable instructions may be provided as utility applications, background daemons, operating system components, or combinations thereof that run in conjunction with processors such as Intel® Xenon® processors or AMD® Opteron® processors and other operating systems known to those skilled in the art, such as Microsoft® Vista®, UNIX®, Solaris®, LINUX®, Apple®, MAC-OS®, and others. Furthermore, the CPU can be implemented as multiple processors that work together in parallel to execute instructions.
[0048] Memory 478 can be a hard disk drive, CD-ROM drive, DVD drive, FLASH® drive, RAM, ROM, or other known electronic storage device in the art.
[0049] Network controllers 474, such as Intel Ethernet® PRO network interface cards from Intel® Corporation, can connect various parts of a PET imaging system. Furthermore, network controllers 474 can also connect to external networks. As can be understood, external networks can be public networks such as the internet, private networks such as LANs and WANs, or any combination thereof, and may include PSTN or ISDN subnetworks. External networks can be wired, such as Ethernet® networks, or wireless, such as cellular networks including EDGE, 3G, and 4G wireless cellular systems. Wireless networks can be Wi-Fi, Bluetooth®, or other known forms of wireless communication.
[0050] The methods and systems described herein can be implemented in many techniques, but generally relate to imaging apparatus and / or processing circuits for performing the processes described herein. In embodiments in which a neural network is used, the processing circuit used to train the neural network does not have to be the same processing circuit used to implement the trained neural network that performs the methods described herein. For example, an FPGA may be used to generate the trained neural network (for example, as defined by its interconnections and weights), and further, a processor 470 and memory 478 may be used to implement the trained neural network. Furthermore, serial or parallel implementations may be used for performance enhancement in the training and use of the trained neural network (for example, by implementing the trained neural network on a parallel processor architecture such as a graphics processor architecture).
[0051] Details have been specifically described in the preceding explanation. However, it should be understood that the technology of this specification may be implemented in other embodiments that deviate from these specific details, and that these details are for illustrative purposes only and not limiting. Embodiments disclosed herein have been described with reference to the accompanying drawings. Similarly, specific figures, materials, and configurations have been shown for illustrative purposes to enhance understanding. However, embodiments may be implemented without these specific details.
[0052] To facilitate understanding of the various embodiments, various techniques have been described as multiple separate operations. The order of description should not be interpreted as indicating that these operations necessarily depend on the order. In fact, these operations do not have to be performed in the order presented. The operations described may be performed in a different order than those in the embodiments described. In additional embodiments, various additional operations may be performed, and / or the operations described may be omitted.
[0053] According to at least one embodiment described above, image quality can be improved.
[0054] While several embodiments have been described, these embodiments are presented as examples only and are not intended to limit the scope of the invention. These embodiments can be implemented in a variety of other forms, and various omissions, substitutions, modifications, and combinations of embodiments are possible without departing from the spirit of the invention. These embodiments and their variations are included in the scope and spirit of the invention, as well as in the claims and their equivalents. [Explanation of symbols]
[0055] 470 Processor 474 Network Controller 476 Data Acquisition System 478 memory
Claims
1. receiving attenuation correction data of a subject and a first nuclear medicine image of the subject, the first nuclear medicine image including scatter effects; and a processing circuit for inputting both the received attenuation correction data and the received first nuclear medicine image into a trained neural network that outputs a scatter sinogram, the processing circuit outputting the scatter sinogram indicative of scatter effects in the first nuclear medicine image; a nuclear medicine imaging system, wherein the trained neural network is trained by inputting as input data sets: (1a) attenuation-corrected training data and (1b) training nuclear medicine images including scatter effects; and (2) a corresponding set of scatter sinogram training outputs having reduced scatter effects compared to the training nuclear medicine images, the scatter sinogram training outputs being generated using a scatter estimation technique different from that of the neural network.
2. The nuclear medicine imaging device according to claim 1 , wherein the trained neural network includes at least one deep convolutional neural network (DCNN).
3. The nuclear medicine imaging device according to claim 1 , wherein the trained neural network includes at least one deep convolutional neural network (DCNN) trained using supervised learning.
4. 2. The nuclear medicine imaging device according to claim 1, wherein the trained neural network includes at least one deep convolutional neural network (DCNN) trained using supervised learning with a similarity measure.
5. The nuclear medicine imaging device according to claim 4 , wherein the similarity measure comprises at least one of a root mean square error, a weighted sum of intensity differences, a cross-correlation, an adversarial loss, and mutual information between image histograms.
6. 2. The nuclear medicine imaging device of claim 1, wherein the processing circuitry further reconstructs a second nuclear medicine image by processing the scatter sinogram indicative of estimated scatter in the first nuclear medicine image.
7. The nuclear medicine imaging device of claim 1 , wherein the set of training scatter sinogram outputs with reduced scatter effects compared to the training nuclear medicine images is based on a corresponding Monte Carlo simulation.
8. The nuclear medicine imaging device of claim 1 , wherein the set of scatter sinogram training outputs with reduced scatter effects compared to the training nuclear medicine images is based on model-based scatter correction.
9. The nuclear medicine imaging device according to claim 1 , wherein the attenuation correction training data is obtained from a computed tomography (CT) scan.
10. The nuclear medicine imaging apparatus of claim 1 , wherein the first nuclear medicine image of the subject is converted to a lower resolution before being input to the trained neural network.
11. The nuclear medicine diagnosis apparatus according to claim 1 , wherein the nuclear medicine diagnosis apparatus is a PET (Positron Emission Tomography) apparatus.
12. The processing circuitry further comprises: (a) generating a difference image from a scatter-corrected image in which scatter is reduced compared to the first nuclear medicine image and (b) the first nuclear medicine image; The nuclear medicine imaging apparatus according to claim 1 , configured to generate a sinogram image by forward projecting the difference image.
13. receiving, by a nuclear medicine imaging device, attenuation correction data of a subject and a first nuclear medicine image of the subject including a scattering effect; An image processing method for outputting a scatter sinogram showing a scattering effect in the first nuclear medicine image by inputting both the received attenuation correction data and the received first nuclear medicine image into a trained neural network that outputs a scatter sinogram, the method comprising: the trained neural network is trained by (1) inputting as input data sets (1a) attenuation-corrected training data and (1b) training nuclear medicine images including scatter effects, and (2) a corresponding set of scatter sinogram training outputs having reduced scatter effects compared to the training nuclear medicine images; The image processing method, wherein the scatter sinogram training output is generated using a scatter estimation technique different from the neural network.
14. On the computer, receiving, by a nuclear medicine imaging device, attenuation correction data of a subject and a first nuclear medicine image of the subject including a scattering effect; A program for causing a nuclear medicine diagnosis device to execute a process of inputting both the received attenuation correction data and the received first nuclear medicine image into a trained neural network that outputs a scatter sinogram, thereby outputting a scatter sinogram that indicates a scattering effect in the first nuclear medicine image, the trained neural network is trained by (1) inputting as input data sets (1a) attenuation-corrected training data and (1b) training nuclear medicine images including scatter effects, and (2) a corresponding set of scatter sinogram training outputs having reduced scatter effects compared to the training nuclear medicine images; The scatter sinogram training output is generated using a scatter estimation technique other than the neural network.