Information processing system, information processing method, and program

Abstract

To enable acquisition of information including at least either of disease name information and disease determination support information related to Parkinson's syndrome on the basis of brain MRI images.SOLUTION: An information processing system is disclosed, comprising an acquisition unit for acquiring input information including 3D data on brain MRI images of a subject, and an output unit configured to use the input information as input data to a machine learning model to output disease name information or disease determination support information related to Parkinson's syndrome for the subject.SELECTED DRAWING: Figure 2

Description

Technical Field

[0001] The present disclosure relates to an information processing system, an information processing method, and a program.

Background Art

[0002] There is known a technique of acquiring one or more MRI image data imaged at one or more time points regarding a subject's brain, and inputting the acquired MRI image data regarding the brain into a learned model that outputs a pseudo-PET image regarding the brain, to output a pseudo-PET image at one or more time points regarding the subject's brain.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, in the prior art as described above, it is difficult to obtain disease name information related to Parkinson's syndrome and disease diagnosis support information (disease duration, cognitive function test results, etc.) based on an MRI image of the brain.

[0005] Therefore, in one aspect, an object of the present disclosure is to make it possible to obtain information including at least one of disease name information related to Parkinson's syndrome and disease diagnosis support information based on an MRI image of the brain.

Means for Solving the Problems

[0006] In one aspect, an acquisition unit that acquires input information including 3D data related to an MRI image of the brain of one subject, An information processing system is provided, which includes an output unit that uses the aforementioned input information as input data and outputs disease name information or disease determination support information related to Parkinson's syndrome for one subject based on a machine learning model. [Effects of the Invention]

[0007] In one respect, this disclosure makes it possible to obtain information that includes at least one of the following: disease name information related to Parkinson's syndrome or disease diagnosis support information, based on MRI images of the brain. [Brief explanation of the drawing]

[0008] [Figure 1] This is a diagram illustrating the overview of this embodiment. [Figure 2] This figure shows some examples of suitable relationships between input data and training data for building a trained model in this embodiment. [Figure 3] This diagram illustrates the types of prediction algorithms suitable for building a trained model in this embodiment. [Figure 4] This is a system diagram showing the overall system configuration of this embodiment. [Figure 5] This figure shows some examples of suitable input data and training data datasets (analysis patterns) in this embodiment. [Figure 6] This is an explanatory diagram of an example of a neural network used in this embodiment. [Figure 7] Figure 5 is an explanatory diagram of a suitable neural network structure for analysis pattern 1. [Figure 8] This figure shows the prediction results (Example 1) obtained using the neural network structure shown in Figure 7 for analysis pattern 1 shown in Figure 5. [Figure 9] This figure shows the prediction results (Example 2) obtained using the neural network structure shown in Figure 7 for analysis pattern 1 shown in Figure 5. [Figure 10] Figure 5 is an explanatory diagram of a suitable neural network structure for analysis pattern 5. [Figure 11] It is a diagram showing the prediction results (Example 3) obtained with the neural network structure shown in FIG. 10 for the analysis pattern 5 shown in FIG. 5. [Figure 12] It is a diagram showing the prediction results (Example 3) obtained with the neural network structure shown in FIG. 10 for the analysis pattern 5 shown in FIG. 5. [Figure 13] It is an explanatory diagram of a suitable neural network structure for the analysis pattern 6 shown in FIG. 5. [Figure 14] It is a diagram showing the prediction results (Example 4) obtained with the neural network structure shown in FIG. 13 for the analysis pattern 6 shown in FIG. 5. [Figure 15] It is a diagram showing the prediction results (Example 4) obtained with the neural network structure shown in FIG. 13 for the analysis pattern 6 shown in FIG. 5. [Figure 16] It is an explanatory diagram of a suitable neural network structure for the analysis pattern 7 shown in FIG. 5. [Figure 17] It is a diagram showing the prediction results (Example 5) obtained with the neural network structure shown in FIG. 16 for the analysis pattern 7 shown in FIG. 5. [Figure 18] It is a diagram showing the prediction results (Example 5) obtained with the neural network structure shown in FIG. 16 for the analysis pattern 7 shown in FIG. 5. [Figure 19] It is an explanatory diagram of a suitable neural network structure for the analysis pattern 8 shown in FIG. 5. [Figure 20] It is a diagram showing the prediction results (Example 6) obtained with the neural network structure shown in FIG. 19 for the analysis pattern 8 shown in FIG. 5. [Figure 21] It is a diagram showing the prediction results (Example 6) obtained with the neural network structure shown in FIG. 19 for the analysis pattern 8 shown in FIG. 5.

Mode for Carrying Out the Invention

[0009] Hereinafter, each embodiment will be described in detail while referring to the accompanying drawings. Note that the dimensional ratios in the drawings are merely examples and are not limited thereto, and the shapes and the like in the drawings may be exaggerated partially for the sake of explanation.

[0010] FIG. 1 is a diagram for explaining the outline of this embodiment.

[0011] As shown in FIG. 1, this embodiment relates to an information processing system that uses a learned model generated by machine learning (for example, deep learning) with clinically significant data (data obtained from clinical and imaging) as teacher data and data obtained from a specific examination modality (image data at a certain time and other additional data) as input data as a prediction analysis algorithm. That is, this embodiment is an information processing system that outputs disease name information or disease diagnosis support information (disease duration, cognitive function test results, etc.) in Parkinson's syndrome using a learned model. The disease diagnosis support information is a concept that includes other information for supporting the determination of the disease name in Parkinson's syndrome (for example, each probability of being a disease of each disease name and information suggesting a specific disease name).

[0012] At this time, the examination modality for obtaining the input data may be not only single but also a combination of a plurality. Metadata such as data acquisition conditions for the input data and the teacher data is also included in the prediction analysis algorithm. For example, imaging conditions, image conditions, protocol conditions such as body position conversion and exercise load can be mentioned.

[0013] First, Parkinson's syndrome is a concept that includes various types of diseases such as Parkinson's disease, progressive supranuclear palsy, and multiple system atrophy. Most of Parkinson's syndrome is a progressive neurological disorder, but while Parkinson's disease has a good treatment response, other Parkinson's syndromes such as progressive supranuclear palsy and multiple system atrophy generally have poor treatment responses, so it is important to distinguish these disease names. However, this distinction is often difficult even for specialists.

[0014] Secondly, in Parkinson's syndrome, neurological symptoms become apparent as the disease progresses. However, in the early stages of Parkinson's syndrome, symptoms progress insidiously, making them difficult to notice not only for the patient and their family but also for doctors. Often, several years have already passed by the time the disease is noticed. At this point, the duration of the disease since the onset of neurological symptoms affects the response to treatment, so it is important to gather information and identify this duration. However, this determination is often difficult. Also, although there are individual differences in the rate of disease progression, it is difficult in the early stages of the disease to determine whether the rate of disease progression in a particular patient is faster or slower than average.

[0015] Thirdly, Parkinson's syndrome is often accompanied by cognitive impairment, and its diagnosis requires cognitive function tests conducted by medical professionals such as specialists, clinical psychologists, and speech-language pathologists. However, repeatedly conducting and evaluating these tests in the routine practice of medical institutions is impractical due to the time and effort involved. On the other hand, delaying the evaluation of cognitive function tests could lead to delays in treatment intervention, raising concerns that symptoms may progress.

[0016] Finally, in clinical trials for Parkinson's syndrome, the determination of disease name, duration of illness, and cognitive impairment is important because they influence the results when evaluating the efficacy of a specific drug within individual patients and when comparing the efficacy of a specific drug among multiple patient groups assigned to that drug. However, as mentioned in points one through three, determining the disease name, duration of illness, and cognitive impairment is not easy.

[0017] Therefore, the information processing system according to this embodiment outputs information regarding at least one of the following in relation to Parkinson's syndrome: disease name, duration of illness, and assessment of cognitive impairment, as disease name information or disease assessment support information related to Parkinson's syndrome. This disease name information and disease assessment support information are intended to be used for drug efficacy evaluation and comparison. Figure 2 shows some examples of the relationship between input data and training data suitable for building a trained model in this embodiment. Figure 3 is an explanatory diagram of the types of prediction algorithms suitable for building a trained model in this embodiment.

[0018] The input data includes 3D MRI image data of the subject's (e.g., patient's) brain.

[0019] Brain-related MRI 3D image data consists of at least one of sagittal and lateral image data, and imaging sequences include 3DT1WI (3DT1-weighted imaging), SWI (Susceptibility-Weighted Imaging), PRESTO (Principle of Echo Shifting with a Train of Observations), T2*-weighted imaging, Neuromelanin MRI, etc.

[0020] Furthermore, the input data may include additional information as appropriate. In this case, the additional information may include the subject's age, gender, Hoehn-Yahr classification, RBDSQ-J (RBD Screening Questionnaire - Japanese Version), etc.

[0021] The training data may include at least one of the following: disease name, duration of illness, and cognitive function test results. In this case, the duration of illness refers to the period from the onset of symptoms to the acquisition of an MRI scan. Cognitive function tests include MoCA-J (Montreal Cognitive Assessment - Japanese version), FAB (Frontal Assessment Battery), TMT-A (Trail Making Test-A), TMT-B (Trail Making Test-B), etc.

[0022] In this case, the trained model is constructed by implementing a convolutional neural network obtained through machine learning. Note that if the training data consists of disease names, the prediction algorithm will be a classification algorithm; if the training data consists of disease duration or cognitive function test results, the prediction algorithm will be a regression algorithm.

[0023] Figure 4 is a system diagram showing the overall schematic configuration of the system in this embodiment.

[0024] As shown in Figure 4, the system 100 includes a learning unit 10, a trained model 11, a storage unit 12, an input unit 13, and a predictive analysis unit 14. Some or all of these units may be implemented by one or more server computers, one or more user terminals (e.g., hospital terminals), or a combination thereof. Furthermore, some of the functions of the learning unit 10, input unit 13, predictive analysis unit 14, etc., can be implemented by the CPU (Central Processing Unit) of a processing unit such as a computer executing a program in a storage device. For example, the processing unit of a user terminal can implement various processes described below by downloading a program from a server. Alternatively, the program can be recorded on a recording medium, and the processing unit can read this recording medium containing the program to implement various processes described below. Various types of recording media can be used. For example, the recording medium may be a recording medium that records information optically, electrically, or magnetically, such as a CD (Compact Disc)-ROM, flexible disk, or magneto-optical disk, or a semiconductor memory that records information electrically, such as ROM or flash memory.

[0025] The input unit 13 receives 3D MRI image data (test MRI data) D15 of the subject's brain.

[0026] The learning unit 10 uses 3D MRI image data D11 related to the brain stored in the memory unit 12 as input data, and various data D12 such as disease names stored in the memory unit 12 as training data, to generate a trained model 11 through deep learning based on a neural network NN. The image data file format may be DICOM, TIFF, JPG, or other formats.

[0027] The memory unit 12 may store a dataset of input data and training data collected from one or more hospitals. For example, the memory unit 12 may store input data D11 of MRI image data (including 3D MRI image data) of the brains of multiple subjects, and training data D12 of the aforementioned disease names etc. related to each subject.

[0028] The predictive analysis unit 14 uses the trained model 11 to predict and analyze the subject's disease name, etc., based on the MRI image 3D data D15 of the subject's brain input from the input unit 13 (step S12). At this time, the subject's age, etc., may be additionally input to take into account factors that may change with age.

[0029] The predictive analysis results (learning results) from the predictive analysis unit 14 are used as disease name information and disease diagnosis support information (step S13).

[0030] A model 10a under training may be generated using a dataset of input data and training data collected from one or more hospitals, with 90% used as the training dataset D13 and the remaining 10% as the validation dataset D14.

[0031] In the future, we may collect datasets of 3D MRI image data D16 and training data D17 related to the brain from one or more other hospitals and further update the trained model 10b.

[0032] In this way, the system 100 can update the trained model 11 by acquiring and accumulating datasets of input data and training data from multiple hospitals, thereby improving the reliability of the disease name information and disease diagnosis support information output from the trained model 11.

[0033] Figure 5 shows some examples of suitable input data and training data datasets (analysis patterns) in this embodiment.

[0034] In the example shown in Figure 5, the input data consists of two patterns: a first pattern consisting solely of MRI 3D image data, and a second pattern which is a combination of MRI 3D image data related to the brain and other additional data. Furthermore, within each of the first and second patterns, the usage patterns of the MRI 3D image data include a first subpattern of a single imaging sequence and a second subpattern of multiple imaging sequences. In other embodiments, subpatterns may be omitted, or three or more subpatterns may be provided. Also, in the example shown in Figure 5, the imaging sequence related to the first subpattern is 3DT1WI, but other imaging sequences may be used. Similarly, in the example shown in Figure 5, the imaging sequence related to the second subpattern is a combination of 3DT1WI and SWI, but other combinations of imaging sequences may be used. Furthermore, combinations of three or more imaging sequences may be used.

[0035] In the example shown in Figure 5, additional data includes age, sex, Hoehn-Yahr classification, and RBDSQ-J, but in other embodiments, some of these may be omitted, or other data may be added.

[0036] Furthermore, in the example shown in Figure 5, the training data includes a third pattern of disease name alone and a fourth pattern of combination of disease duration and cognitive function test results. In the example shown in Figure 5, the disease names are three types: Parkinson's disease (PD), progressive supranuclear palsy (PSP), and multiple system atrophy (MSA). However, in other examples, some of these may be omitted, or other types may be added. Also, the cognitive function tests are the four tests mentioned above (i.e., MoCA-J, FAB, TMT-A, and TMT-B). However, in other examples, some of these may be omitted, or other tests may be added.

[0037] In this case, eight analysis patterns as shown in Figure 5 are suitable. In the table shown in Figure 5, the input data and analysis data used in each analysis pattern are marked with black circles. For example, in analysis pattern 1, the input data is pattern 1, and the usage pattern of the MRI 3D image data is sub-pattern 1. The training data is pattern 3. For example, in analysis pattern 8, the input data is pattern 2, and the usage pattern of the MRI 3D image data is sub-pattern 2. The training data is pattern 4.

[0038] These multiple analysis patterns may be used individually or in combination as appropriate. For example, in the system 100 shown in Figure 4, a trained model 11 may be built for each analysis pattern. In this case, one or more of the multiple trained models 11 may be selected and used for a particular subject. The selection of which trained model 11 is chosen may be determined as appropriate based on the attributes of the disease name information or disease diagnosis support information required as output, the attributes of the subject in question, etc.

[0039] Figure 6 is an explanatory diagram of an example of a neural network used in this embodiment. In Figure 6, the 3D Convolution Layer represents a three-dimensional convolutional layer. The 3D Batch Normalization Layer represents a three-dimensional normalization processing layer. The ReLU Layer represents a normalization linear function and is a type of activation function. The 3D ConvBNActiBlock is a block of layers that combine the 3D Convolutional Layer, 3D Batch Normalization, and ReLU Layer in that order.

[0040] Figure 7 is an explanatory diagram of a suitable neural network structure for analysis pattern 1 shown in Figure 5.

[0041] Figure 7 shows an example of an analysis method that uses machine learning to classify disease names (PD, PSP, MSA) from 3DT1WI data of MRI images related to the brain.

[0042] In the example shown in Figure 7, 3DT1WI data is first input, and the output results are input to the fully connected layer, followed by the Dropout layer. Then, the data is input to the final fully connected layer to output the resulting disease name (PD, PSP, MSA).

[0043] Specifically, the first 3D ConvBNActiBlock has a kernel size of 3x3x3 (depth x height x width), a stride of 2x2x2 (depth x height x width), and no padding. The second 3D ConvBNActiBlock converts from 1x150x256x256 to 64x74x127x127. Next, the second 3D ConvBNActiBlock has a kernel size of 3x3x3, a stride of 2x2x2, and no padding. The second 3D ConvBNActiBlock converts from 64x74x127x127 to 128x36x63x63. Next, the third 3D ConvBNActiBlock has a kernel size of 5x5x5, a stride of 3x3x3, and no padding. The third 3D ConvBNActiBlock converts from 128x36x63x63 to 512x11x20x20. Next, the fourth 3D ConvBNActiBlock has a kernel size of 5x5x5, a stride of 3x3x3, and no padding. The fourth 3D ConvBNActiBlock converts from 512x1x20x20 to 512x3x6x6. Then, one 3D Convolution Layer (kernel size 1x6x6, stride 3x3x3, no padding) converts from 512x3x6x6 to 512x1x1x1. Finally, flatten it to a 1x512 vector. Then, a 512x16 fully connected layer converts from 1x512 to 1x16, and the final 16x3 fully connected layer converts from 1x16 to 1x3, which is then input to a sigmoid layer, outputting PD, PSP, and MSA.

[0044] Figure 8 shows the prediction results (Example 1) performed using the neural network structure shown in Figure 7 for analysis pattern 1 shown in Figure 5. Figure 9 shows the prediction results (Example 2) performed using the neural network structure shown in Figure 7 for analysis pattern 1 shown in Figure 5.

[0045] Regarding Examples 1 and 2 shown in Figures 8 and 9, Adam was used as the optimizer to construct the neural network structure shown in Figure 7. During training, the learning rate was set to 0.7 times the previous learning rate every 15 epochs. The Binary Cross Entropy (BCE) loss function was used. The batch size for training and validation (test) data was 1, the learning rate was 0.0001, the development environment was Windows 10 (registered trademark), and the framework used was PyTorch (registered trademark).

[0046] In Figures 8 and 9, the ROC curve of PD is the ROC (Receiver Operating Characteristic) curve used to distinguish between PD and others, the ROC curve of PSP is the ROC curve used to distinguish between PSP and others, the ROC curve of MSA is the ROC curve used to distinguish between MSA and others, and the micro-average ROC curve is the ROC curve used to distinguish between all three.

[0047] Figures 8 and 9 show that PD, PSP, and MSA can be classified from 3DT1WI data of MRI images of the brain using machine learning, confirming the effectiveness of the neural network structure shown in Figure 7.

[0048] Figure 10 is an explanatory diagram of a suitable neural network structure for analysis pattern 5 shown in Figure 5.

[0049] Figure 10 shows an example of an analysis method that uses machine learning to predict disease duration and cognitive function test results from 3DT1WI data of MRI images related to the brain.

[0050] In the example shown in Figure 10, the 3DT1WI data is first input to Branch 1, the output of Branch 1 is input to the fully connected layer, and then to the Dropout layer. Finally, it is input to the last fully connected layer to output the results: disease duration, MoCA-J, FAB, TMT-A, and TMT-B.

[0051] Specifically, Branch1 consists of four 3DConvBNActi Blocks and one 3D Convolution Layer. The first 3D ConvBNActiBlock has a kernel size of 3x3x3 (depth x height x width), a stride of 2x2x2 (depth x height x width), and no padding. The first 3D ConvBNActiBlock converts from 1x176x256x256 to 64x87x127x127. The second 3D ConvBNActiBlock has a kernel size of 3x3x3, a stride of 2x2x2, and no padding. The second 3D ConvBNActiBlock converts from 64x87x127x127 to 128x43x63x63. The third 3D ConvBNActiBlock has a kernel size of 5x5x5, a stride of 3x3x3, and no padding. The third 3D ConvBNActiBlock converts from 128x43x63x63 to 512x13x20x20. The fourth 3D ConvBNActiBlock has a kernel size of 5x5x5, a stride of 3x3x3, and no padding. The fourth 3D ConvBNActiBlock converts from 512x13x20x20 to 512x3x6x6. One 3D Convolution Layer has a kernel size of 1x6x6, a stride of 3x3x3, and no padding. One 3D Convolution Layer converts from 512x3x6x6 to 512x1x1x1. Then, it is flattened to convert it into a 1x512 vector. Then, a 512×16 fully connected layer converts from 1×512 to 1×16, and the final 16×5 fully connected layer converts from 1×16 to 1×5, outputting disease duration, MoCA-J, FAB, TMT-A, and TMT-B.

[0052] Figures 11 and 12 show the prediction results (Example 3) performed using the neural network structure shown in Figure 10 for analysis pattern 5 shown in Figure 5.

[0053] Regarding Example 3 shown in Figures 11 and 12, Adam was used as the optimizer to construct the neural network structure shown in Figure 10. During training, the learning rate was set to 0.7 times the previous learning rate every 15 epochs. The mean squared error (MSE) loss function was used. The batch size of the training and validation (test) data was 1, the learning rate was 0.0001, the development environment was Windows 10®, and the framework used was PyTorch®.

[0054] Figures 11 and 12 show the prediction results (MSE) for 10 cases. In this case, 4 of the 10 cases were used as training data and 6 as validation (test) data. Predicted values ​​for disease duration, MoCA-J, FAB, TMT-A, and TMT-B were calculated, and scatter plots in Figures 11 and 12 were obtained showing the relationship between each predicted value and the true value. Figure 11 shows the scatter plots 1101, 1102, and 1103 for disease duration, MoCA-J, and FAB, respectively, and Figure 12 shows the scatter plots 1104 and 1105 for TMT-A and TMT-B, respectively.

[0055] Figures 11 and 12 show that disease duration and cognitive function test results (MoCA-J, FAB, TMT-A, and TMT-B) can be predicted using machine learning from 3DT1WI data of MRI images of the brain, confirming the effectiveness of the neural network structure shown in Figure 10.

[0056] Figure 13 is an explanatory diagram of a suitable neural network structure for analysis pattern 6 shown in Figure 5.

[0057] Figure 13 shows an example of an analysis method that uses machine learning to predict disease duration and cognitive function test results from 3DT1WI and SWI data of MRI images related to the brain.

[0058] In the example shown in Figure 13, first, 3DT1WI data is input to Branch1, and SWI data is input to Branch2. The output results of Branch1 and Branch2 are added together, and the output result is input to a fully connected layer, and then to a Dropout layer. Finally, it is input to the last fully connected layer, and the result is output.

[0059] Specifically, both Branch1 and Branch2 consist of four 3DConvBNActi Blocks and one 3D Convolution Layer.

[0060] In Branch 1, the first 3D ConvBNActiBlock has a kernel size of 3x3x3 (depth x height x width), a stride of 2x2x2 (depth x height x width), and no padding. The first 3D ConvBNActiBlock converts from 1x176x256x256 to 64x87x127x127. The second 3D ConvBNActiBlock has a kernel size of 3x3x3, a stride of 2x2x2, and no padding. The second 3D ConvBNActiBlock converts from 64x87x127x127 to 128x43x63x63. The third 3D ConvBNActiBlock has a kernel size of 5x5x5, a stride of 3x3x3, and no padding. The third 3D ConvBNActiBlock converts from 128×43×63×63 to 512×13×20×20. The fourth 3D ConvBNActiBlock has a kernel size of 5×5×5, a stride of 3×3×3, and no padding. The fourth 3D ConvBNActiBlock converts from 512×13×20×20 to 512×3×6×6. Next, one 3D Convolution Layer converts from 512×3×6×6 to 512×1×1×1. This 3D Convolution Layer has a kernel size of 1×6×6, a stride of 3×3×3, and no padding.

[0061] In Branch 2, the first 3D ConvBNActiBlock has a kernel size of 3x3x3 (depth x height x width), a stride of 2x2x2 (depth x height x width), and no padding. The first 3D ConvBNActiBlock converts from 1x89x290x320 to 64x44x144x159. The second 3D ConvBNActiBlock has a kernel size of 3x3x3, a stride of 2x2x2, and no padding. The second 3D ConvBNActiBlock converts from 64x44x144x159 to 128x21x71x79. The third 3D ConvBNActiBlock has a kernel size of 5x5x5, a stride of 3x3x3, and no padding. The third 3D ConvBNActiBlock converts from 128×21×71×79 to 512×6×23×25. The fourth 3D ConvBNActiBlock has a kernel size of 5×5×5, a stride of 3×3×3, and no padding. The fourth 3D ConvBNActiBlock converts from 512×6×23×25 to 512×1×7×7. Next, one 3D Convolution Layer converts from 512×1×7×7 to 512×1×1×1. The 3D Convolution Layer has a kernel size of 1×6×6, a stride of 3×3×3, and no padding.

[0062] In this case, the output results of Branch1 and Branch2 are added together, flattened into a 1×512 vector, and then converted from 1×512 to 1×16 in a 512×16 fully connected layer. Finally, in the last 16×5 fully connected layer, it is converted from 1×16 to 1×5, and the disease duration, MoCA-J, FAB, TMT-A, and TMT-B are output.

[0063] Figures 14 and 15 show the prediction results (Example 4) performed using the neural network structure shown in Figure 13 for analysis pattern 6 shown in Figure 5.

[0064] Regarding Example 4 shown in Figures 14 and 15, Adam was used as the optimizer to construct the neural network structure shown in Figure 13. During training, the learning rate was set to 0.7 times the previous learning rate every 15 epochs. The mean squared error (MSE) loss function was used. The batch size of the training and validation (test) data was 1, the learning rate was 0.0001, the development environment was Windows 10®, and the framework used was PyTorch®.

[0065] Figures 14 and 15 show the prediction results (MSE) for 10 cases. In this case, 4 of the 10 cases were used as training data and 6 as validation (test) data. Predicted values ​​for disease duration, MoCA-J, FAB, TMT-A, and TMT-B were calculated, and scatter plots in Figures 14 and 15 were obtained showing the relationship between each predicted value and the true value. Figure 14 shows the scatter plots 1401, 1402, and 1403 for disease duration, MoCA-J, and FAB, respectively, and Figure 15 shows the scatter plots 1404 and 1405 for TMT-A and TMT-B, respectively.

[0066] Figures 14 and 15 show that the duration of the disease and the results of cognitive function tests (MoCA-J, FAB, TMT-A, and TMT-B) can be predicted using machine learning from 3DT1WI and SWI data of MRI images of the brain, confirming the effectiveness of the neural network structure shown in Figure 13.

[0067] Figure 16 is an explanatory diagram of a suitable neural network structure for analysis pattern 7 shown in Figure 5.

[0068] Figure 16 shows an example of an analysis method that uses machine learning to predict disease duration and cognitive function test results from 3DT1WI data of brain MRI images and other information (age, sex, Hoehn-Yahr classification, RBDSQ-J).

[0069] In the example shown in Figure 16, 3DT1WI data is first input to Branch1, the output of Branch1 is input to a fully connected layer, and then to a Dropout layer. After that, the result is joined with age, sex, Hoehn-Yahr classification, and RBDSQ-J, and then input to the final fully connected layer to output the result.

[0070] Specifically, Branch1 consists of four 3DConvBNActi Blocks and one 3D Convolution Layer. The first 3D ConvBNActiBlock has a kernel size of 3x3x3 (depth x height x width), a stride of 2x2x2 (depth x height x width), and no padding. The first 3D ConvBNActiBlock converts from 1x176x256x256 to 64x87x127x127. The second 3D ConvBNActiBlock has a kernel size of 3x3x3, a stride of 2x2x2, and no padding. The second 3D ConvBNActiBlock converts from 64x87x127x127 to 128x43x63x63. The third 3D ConvBNActiBlock has a kernel size of 5x5x5, a stride of 3x3x3, and no padding. The third 3D ConvBNActiBlock converts from 128x43x63x63 to 512x13x20x20. The fourth 3D ConvBNActiBlock has a kernel size of 5x5x5, a stride of 3x3x3, and no padding. The fourth 3D ConvBNActiBlock converts from 512x13x20x20 to 512x3x6x6. Next, one 3D Convolution Layer converts from 512x3x6x6 to 512x1x1x1. The 3D Convolution Layer has a kernel size of 1x6x6, a stride of 3x3x3, and no padding. Next, the data is flattened to convert it into a 1×512 vector, and then converted from 1×512 to 1×16 in a 512×16 fully connected layer. Finally, in a 16×5 fully connected layer, it is joined with age, sex, Hoehn-Yahr classification, and RBDSQ-J, then converted from 1×16 to 1×5 to output disease duration, MoCA-J, FAB, TMT-A, and TMT-B.

[0071] Figures 17 and 18 show the prediction results (Example 5) obtained by performing the analysis pattern 7 shown in Figure 5 using the neural network structure shown in Figure 16.

[0072] Regarding Example 5 shown in Figures 17 and 18, Adam was used as the optimizer to construct the neural network structure shown in Figure 16. During training, the learning rate was set to 0.7 times the previous learning rate every 15 epochs. The mean squared error (MSE) loss function was used. The batch size of the training and validation (test) data was 1, the learning rate was 0.0001, the development environment was Windows 10®, and the framework used was PyTorch®.

[0073] Figures 17 and 18 show the prediction results (MSE) for 10 cases. In this case, 4 of the 10 cases were used as training data and 6 as validation (test) data. Predicted values ​​for disease duration, MoCA-J, FAB, TMT-A, and TMT-B were calculated, and scatter plots in Figures 17 and 18 were obtained showing the relationship between each predicted value and the true value. Figure 17 shows scatter plots 1701, 1702, and 1703 for disease duration, MoCA-J, and FAB, respectively, and Figure 18 shows scatter plots 1704 and 1705 for TMT-A and TMT-B, respectively.

[0074] Figures 17 and 18 show that the duration of illness and the results of cognitive function tests (MoCA-J, FAB, TMT-A, and TMT-B) can be predicted using machine learning based on 3DT1WI data of brain MRI images and other information (age, sex, Hoehn-Yahr classification, RBDSQ-J), confirming the effectiveness of the neural network structure shown in Figure 16.

[0075] Figure 19 is an explanatory diagram of a suitable neural network structure for analysis pattern 8 shown in Figure 5.

[0076] Figure 19 shows an example of an analysis method that uses machine learning to predict disease duration and cognitive function test results from 3DT1WI and SWI data of brain MRI images and other information (age, sex, Hoehn-Yahr classification, RBDSQ-J).

[0077] In the example shown in Figure 19, 3DT1WI data is first input to Branch1, SWI data to Branch2, the output results of Branch1 and Branch2 are added together, and the output result is input to a fully connected layer, and then to a Dropout layer. Then, the result is joined with age, sex, Hoehn-Yahr classification, and RBDSQ-J, and finally input to the last fully connected layer to output the result.

[0078] Specifically, both Branch1 and Branch2 consist of four 3DConvBNActi Blocks and one 3D Convolution Layer.

[0079] The configurations of Branch1 and Branch2 are the same as in Figure 13 above, so we will omit the explanation. In this case, the output results of Branch1 and Branch2 are added together, flattened to convert them into a 1×512 vector, and then converted from 1×512 to 1×16 in a 512×16 fully connected layer. Finally, in the 16×5 fully connected layer, the data is joined with age, sex, Hoehn-Yahr classification, and RBDSQ-J, then converted from 1×16 to 1×5 to output disease duration, MoCA-J, FAB, TMT-A, and TMT-B.

[0080] Figures 20 and 21 show the prediction results (Example 6) performed using the neural network structure shown in Figure 19 for analysis pattern 8 shown in Figure 5.

[0081] Regarding Example 6 shown in Figures 20 and 21, Adam was used as the optimizer to construct the neural network structure shown in Figure 19. During training, the learning rate was set to 0.7 times the previous learning rate every 15 epochs. The mean squared error (MSE) loss function was used. The batch size of the training and validation (test) data was 1, the learning rate was 0.0001, the development environment was Windows 10®, and the framework used was PyTorch®.

[0082] Figures 20 and 21 show the prediction results (MSE) for 11 cases. In this case, 5 of the 10 cases were used as training data and 6 as validation (test) data. Predicted values ​​for disease duration, MoCA-J, FAB, TMT-A, and TMT-B were calculated, and scatter plots showing the relationship between each predicted value and the true value were obtained in Figures 20 and 21. Figure 20 shows the scatter plots 2001, 2002, and 2003 for disease duration, MoCA-J, and FAB, respectively, and Figure 21 shows the scatter plots 2004 and 2005 for TMT-A and TMT-B, respectively.

[0083] Figures 20 and 21 show that the duration of illness and the results of cognitive function tests (MoCA-J, FAB, TMT-A, and TMT-B) can be predicted using machine learning based on 3DT1WI and SWI data from brain MRI images and other information (age, sex, Hoehn-Yahr classification, RBDSQ-J), confirming the effectiveness of the neural network structure shown in Figure 19.

[0084] Although each embodiment has been described in detail above, the invention is not limited to any particular embodiment, and various modifications and changes are possible within the scope described in the claims. Furthermore, it is possible to combine all or more of the components of the embodiments described above. [Explanation of Symbols]

[0085] 100 Systems 10. Learning Department 11. Pre-trained models 12 Storage section 13 Input section 14. Predictive Analysis Department 15 Display section 16 Memory section

Claims

1. An acquisition unit that acquires input information including 3D data relating to an MRI image of the brain of one subject, The system includes an output unit that uses the aforementioned input information as input data and outputs disease determination support information related to Parkinson's syndrome for one subject based on a machine learning model, The aforementioned disease diagnosis support information includes at least one of the duration of illness and the results of cognitive function tests, in an information processing system.

2. The information processing system according to claim 1, wherein the MRI image includes at least one of a sagittal image of the head and a horizontal image of the head.

3. The information processing system according to claim 1, wherein the output unit further outputs disease name information relating to Parkinson's syndrome for the subject.

4. The information processing system according to claim 1, wherein the input information further includes at least one of age, gender, Hoehn-Yahr classification, and RBDSQ-J.

5. Input information including 3D data related to MRI images of one subject's brain is acquired. Using the aforementioned input information as input data, a machine learning model outputs disease diagnosis support information related to Parkinson's syndrome for the aforementioned subject. Let the computer perform the process, The disease diagnosis support information includes at least one of the following: duration of illness and results of cognitive function tests.

6. Input information including 3D data related to MRI images of one subject's brain is acquired. The system includes outputting disease assessment support information related to Parkinson's syndrome for one subject, based on a machine learning model, using the aforementioned input information as input data. The aforementioned disease diagnosis support information includes at least one of the duration of illness and the results of cognitive function tests, and is an information processing method performed by a computer.

Images