Learning device, trained model generation method, diagnostic processing device, computer program, and diagnostic processing method

Abstract

Provided is a diagnosis processing device that reduces the burden on patients and hospitals.　This diagnosis processing device 100 comprises: a storage unit 12 which stores image data of brain MRI examinations obtained by imaging the brains of a plurality of patients, and a specific binding ratio (SBR) value of brain dopamine transporter scintigraphy examinations corresponding to each of the plurality of patients; an input unit 13 which inputs the image data of the brain MRI examinations of the patients; and a prediction analysis unit 14 which receives, as input data, the image data of the brain MRI examinations of the plurality of patients stored in the storage unit 12, and uses a trained model 11 generated from deep learning by taking, as training data, the SRB value of the brain dopamine transporter scintigraphy examination corresponding to each of the plurality of patients stored in the storage unit 12 to predict and analyze the SBR values of the patients on the basis of the image data of the brain MRI examinations of the patients input from the input unit 13.

Description

[Technical Field]

[0001] The present disclosure relates to a learning device, a method for generating a trained model, a diagnostic processing device, a computer program, and a diagnostic processing method. [Background technology]

[0002] Isotope testing is a technique in which a diagnostic agent containing a radioisotope is injected intravenously to visualize the physiological function of a target area. This technology makes it possible to determine whether physiological functions are normal or abnormal, which cannot be obtained through conventional imaging tests.

[0003] The disease differentiation support device of Patent Document 1 uses image data from a nuclear medicine test that images the head of a subject to whom a dopamine transporter imaging agent has been administered, to assist in differentiating between a first disease, both of which involve degeneration of dopamine neurons in the substantia nigra-striatum, and a second disease that is different from the first disease. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent No. 6549521 Summary of the Invention [Problem to be solved by the invention]

[0005] However, isotope testing has several problems. First, hospitals must order specialized radiological testing agents every time a patient undergoes an examination, which requires complicated and time-consuming management. Second, specialized testing staff such as radiologists are required, and each examination takes a long time, making it labor-intensive. Third, not only the patient undergoing the examination but also the testing staff are at risk of radiation exposure. Fourth, there are few hospitals that can perform this test, and it can only be performed at a limited number of large hospitals. Fifth, the cost per examination is high, which places a heavy financial burden on patients.

[0006] Therefore, an object of the present invention is to provide a learning device, a trained model generation method, a diagnostic processing device, a computer program, and a diagnostic processing method that place less of a burden on patients and hospitals. Specifically, the present invention uses deep learning to predict and analyze the degree of degeneration of dopamine neurons in the substantia nigra-striatum, which corresponds to the SBR (Specific Binding Ratio) value of a brain dopamine transporter scintigraphy test, from image data obtained from a head MRI test. [Means for solving the problem]

[0007] In order to solve the above problem, the learning device of the present invention includes a memory unit that stores image data from head MRI examinations of multiple patients' heads and SBR values ​​from cerebral dopamine transporter scintigraphy examinations corresponding to each patient, and a learning unit that uses the image data from the head MRI examinations of multiple patients stored in the memory unit as input data and the SBR values ​​from the cerebral dopamine transporter scintigraphy examinations corresponding to each patient stored in the memory unit as training data to generate a trained model using deep learning.

[0008] The trained model includes a first neural network that uses right brain images to predict and analyze SBR values ​​of the right brain, and a second neural network that uses left brain images to predict and analyze SBR values ​​of the left brain.

[0009] The trained model includes a third neural network that uses right and left brain images to predict and analyze right and left brain SBR values.

[0010] The trained model includes a fourth neural network that uses a mirrored right or left brain image to predict and analyze the SBR values ​​of the right brain and the left brain.

[0011] The method for generating a trained model of the present invention includes a storage step for storing image data from head MRI examinations of multiple patients' heads and SBR values ​​from cerebral dopamine transporter scintigraphy examinations corresponding to each patient, and a step for generating a model using deep learning with the image data from the head MRI examinations of multiple patients stored in the storage step as input data and the SBR values ​​from the cerebral dopamine transporter scintigraphy examinations corresponding to each patient stored in the storage unit as training data.

[0012] The diagnostic processing device of the present invention has an input unit that inputs image data from a patient's head MRI examination, and a predictive analysis unit that uses image data from head MRI examinations of multiple patients as input data and predicts and analyzes the SBR value of the patient's cerebral dopamine transporter scintigraphy examination based on the image data of the patient's head MRI examination input from the input unit, using a trained model generated by deep learning with the SBR value of the cerebral dopamine transporter scintigraphy examination corresponding to each patient as training data.

[0013] The trained model includes a first neural network that uses right brain images to predict and analyze SBR values ​​of the right brain, and a second neural network that uses left brain images to predict and analyze SBR values ​​of the left brain.

[0014] The trained model includes a third neural network that uses right and left brain images to predict and analyze right and left brain SBR values.

[0015] The trained model includes a fourth neural network that uses a mirrored right or left brain image to predict and analyze the SBR values ​​of the right brain and the left brain.

[0016] The device has a display unit that displays the normal range of the SBR and the SBR value obtained by the predictive analysis.

[0017] The present invention is a computer program to be executed by a diagnostic processing device, which causes the diagnostic processing device to execute a predictive analysis step in which image data from head MRI examinations of multiple patients is used as input data, and the SBR values ​​of the brain dopamine transporter scintigraphy examinations corresponding to each patient are used as training data to predict and analyze the SBR values ​​of the brain dopamine transporter scintigraphy examinations of input patients based on the image data of the head MRI examinations of the input patients.

[0018] The present invention is a diagnostic processing method performed by a diagnostic processing device, which includes an input step of inputting image data from a head MRI examination of a patient's head, and a predictive analysis step of predicting and analyzing the SBR value of a cerebral dopamine transporter scintigraphy examination of a patient based on the image data of the head MRI examination of the patient input in the input step, using image data from the head MRI examinations of a plurality of patients as input data and a trained model generated using deep learning with the SBR value of the cerebral dopamine transporter scintigraphy examination corresponding to each patient as training data. [Effects of the Invention]

[0019] According to the present disclosure, it is possible to provide a learning device, a method for generating a trained model, a diagnostic processing device, a computer program, and a diagnostic processing method that place less burden on patients and medical institutions. [Brief explanation of the drawings]

[0020] [Figure 1] FIG. 1 is a diagram for explaining an overview of the present invention. [Figure 2]This is an example of predictive analysis of the results of brain dopamine transporter scintigraphy from image data from a head MRI examination. [Figure 3] (a) shows a head MRI examination, (b) shows cerebral dopamine transporter scintigraphy, (c) shows input data, and (d) shows training data. [Figure 4] This is an example of predicting and analyzing the SBR value of a cerebral dopamine transporter scintigraphy test from image data of a head MRI test. [Figure 5] FIG. 1 is an explanatory diagram illustrating specific preprocessing of image data from a head MRI examination. [Figure 6] This is a schematic diagram showing the process of resizing and halving the 10 middle DICOM images contained in a sample of image data from a head MRI examination. [Figure 7] FIG. 10 is a diagram illustrating normalization processing of signal intensity between samples. [Figure 8] This is a table showing the average error and mean square error for each of the three models created for each of the neural networks NN1 to NN4. [Figure 9] FIG. 1 is an explanatory diagram of a neural network used in an embodiment of the present invention. [Figure 10] FIG. 10 is a diagram illustrating the results of SBR prediction analysis of the right brain when normalization processing is not performed in the neural network NN1. [Figure 11] FIG. 10 is a diagram illustrating the results of SBR prediction analysis of the right brain when normalization processing is performed in the neural network NN1. [Figure 12] FIG. 10 is a diagram illustrating the results of SBR prediction analysis of the left brain when no normalization processing is performed in the neural network NN2. [Figure 13] FIG. 10 is a diagram illustrating the results of SBR prediction analysis of the left brain when normalization processing is performed in the neural network NN2. [Figure 14] FIG. 10 is a diagram illustrating the results of SBR prediction analysis of the right and left brain when no normalization processing is performed in the neural network NN3. [Figure 15]FIG. 10 is a diagram illustrating the results of SBR prediction analysis of the right and left brain when normalization processing is performed in the neural network NN3. [Figure 16] FIG. 10 is a diagram illustrating the results of SBR prediction analysis of the right and left brain when no normalization processing is performed in the neural network NN4. [Figure 17] FIG. 10 is a diagram illustrating the results of SBR prediction analysis of the right and left brain when normalization processing is performed in the neural network NN4. [Figure 18] This is an example screen from software that predicts and analyzes the SBR value of a cerebral dopamine transporter scintigraphy test from image data from a head MRI test. The screen shows how to select a folder containing DICOM image files to be analyzed and import all files within it. [Figure 19] This is an example of a screen displaying all DICOM images contained in a folder. [Figure 20] 10 is an example of a display screen of an SBR predicted analysis value displayed on a display unit of the diagnostic processing device in the embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0021] Hereinafter, embodiments of the present disclosure will be described in detail.

[0022] FIG. 1 is a diagram for explaining an overview of the present invention. As shown in Figure 1, this invention is software that incorporates, as a predictive analysis algorithm, a trained model generated using machine learning or deep learning with quantitative values ​​obtained from isotope testing as training data and data obtained from a specific testing modality (image data or time-series quantitative data) as input data. This invention aims to limit the patients who truly need isotope testing or to replace isotope testing by obtaining predictive analysis results comparable to those of isotope testing from a less invasive testing modality.

[0023] In this case, the test modality used to obtain the input data may be a single one or a combination of multiple modalities. Metadata such as the data acquisition conditions for the input data and training data are also included in the predictive analysis algorithm. Examples include imaging conditions, image conditions, and protocol conditions such as postural changes and exercise load.

[0024] Figure 2 shows an example of predictive analysis of the results of brain dopamine transporter scintigraphy from image data from a head MRI examination. Brain dopamine transporter scintigraphy is an imaging test in which radioactive iodine-labeled isoflurane is injected intravenously, and the isoflurane bound to the dopamine transporter (DAT) in the dopamine nerve endings of the substantia nigra-striatum is imaged with a gamma camera to visualize the density and distribution of the DAT.

[0025] Brain dopamine transporter scintigraphy is a test performed on patients suspected of having Parkinson's syndrome, and is used to differentiate essential tremor and drug-induced parkinsonism from other parkinsonian syndromes. While the test results are normal in essential tremor and drug-induced parkinsonism, they are abnormal in other parkinsonian syndromes. The SBR value is output as the quantitative value. To calculate the SBR value, a region of interest (ROI) is set in the striatum and the whole brain excluding the striatum. The SBR value is then calculated from the total count in the striatum region, the count per unit volume in the whole brain excluding the striatum, the volume of the striatum region, and the actual volume of the striatum. This reflects the distribution, density, and function of dopamine-producing neurons, and allows the degree of degeneration and loss to be determined.

[0026] We will create a neural network that predicts and analyzes the SBR value of a brain dopamine transporter scintigraphy test from image data of a head MRI test. The imaging sequences for head MRI examinations include T1 weighted images, T2 weighted images, FLAIR images (fluid-attenuated inversion recovery), diffusion weighted images, and T2* These include images, SWI (susceptibility-weighted imaging) images, melanin-enhanced images, and MRA (MR-angiography) images. Image data file formats include DICOM, TIFF, JPG, etc.

[0027] Imaging conditions include the MRI manufacturer name, model name, magnetic field strength, coil name (number of channels), repetition time (TR) for each sequence, echo time (TE), inversion time (TI), slice thickness, field of view, matrix size, parallel imaging / acceleration factor, band width, number of excitations, scan time, etc. The SBR value (teaching data) for cerebral dopamine transporter scintigraphy tests uses a standardized value between facilities. Imaging conditions include the gamma camera manufacturer name, model name, collimator name, drug manufacturer name, dosage at test time, actual dosage, energy, matrix size, magnification rate, time from administration to imaging, and time from start to finish. If the data collection mode is continuous rotation collection mode, it also includes the number of seconds per phase, rotation angle, number of views, etc. If the data collection mode is step and shoot mode, it also includes the step angle, number of seconds per step, step and shoot time, etc.

[0028] Figure 3(a) shows a head MRI examination, (b) cerebral dopamine transporter scintigraphy, (c) input data, and (d) training data. All image data (eg, DICOM images) from the cranial to caudal regions in one type of imaging sequence (eg, FLAIR images) obtained in one head MRI examination of one patient is defined as one sample. In this study, we developed a neural network using samples from multiple patients as input data and the brain dopamine transporter scintigraphy data (Right SBR value / Left SBR value) corresponding to each patient as training data.

[0029] FIG. 4 shows an example of predicting and analyzing the SBR value of a patient's brain dopamine transporter scintigraphy test from image data of the patient's head MRI test. As shown in FIG. 4, the diagnostic processing device 100 includes a learning unit 10, a trained model 11, a memory unit 12, an input unit 13, a predictive analysis unit 14, a display unit 15, and a memory unit 16.

[0030] The input unit 13 inputs image data (test MRI data) D15 from a head MRI examination of a patient. The learning unit 10 uses the head MRIDICOM data D11 stored in the memory unit 12 as input data and the SBR value D12 of the patient's cerebral dopamine transporter scintigraphy test stored in the memory unit 12 as training data to generate a trained model 11 through deep learning. In addition, when generating the trained model 11, the input data may include image data of the head MRI tests of multiple patients, and training data such as metadata on the data acquisition conditions of the cerebral dopamine transporter scintigraphy test corresponding to each patient.

[0031] The trained model 11 includes a first neural network NN1, a second neural network NN2, a third neural network NN3, and a fourth neural network NN4, which will be described later. The predictive analysis unit 14 uses the trained model 11 to predict and analyze the SBR value of the patient based on the image data D15 of the patient's head MRI examination input from the input unit 13 (step S12). At this time, the patient's age is also input for comparison with the normal range of SBR, which changes with age. The display unit 15 displays various screens of the diagnostic processing device 100 based on the data stored in the storage unit 16, and also displays the normal range of SBR and the SBR value obtained by predictive analysis.

[0032] Input data and training data were collected from one or more hospitals, and a supervised learning neural network (NN) was generated using these datasets. 90% of the dataset was used as the training dataset (D13), and the remaining 10% was used as the validation dataset (D14). While model learning was continued using the training data set D13 and the validation data set D14 with the model 10a under training, the already trained trained model 10 was used to predict and analyze SBR in the prediction analysis unit 14 from test MRI data D15 input to the input unit 13 (step S12), and the model learning results were evaluated (step S13). In this case, the test MRI data D15 was substituted with the same data as the validation MRI data D14.

[0033] In the future, we will collect data sets of head MRIDICOM data D16 and dopamine transporter scintigraphy data D17 from several other hospitals and further update the trained model 10b.

[0034] Among the components of the diagnostic processing device 100, the storage unit 12 stores input data, such as image data (head MRIDICOM data) D11 from head MRI examinations of multiple patients, and training data, such as SBR values ​​(dopamine transporter scintigraphy data) D12 from cerebral dopamine transporter scintigraphy examinations corresponding to each patient. The storage unit 12 also stores input data, such as image data from head MRI examinations of multiple patients, and training data, such as metadata on data acquisition conditions for cerebral dopamine transporter scintigraphy examinations corresponding to each patient.

[0035] FIG. 5 is an explanatory diagram showing specific pre-processing of image data from a head MRI examination. Figure 6 shows a schematic diagram of the process of resizing and halving the 10 middle DICOM images contained in a sample of image data from a head MRI examination. One sample refers to all image data (e.g., DICOM images) from the cranial to caudal regions included in one imaging sequence (e.g., FLAIR images) obtained from one head MRI examination of one patient (step S21). For multiple patients, the middle 10 images within the sample were selected, and these 10 DICOM images were resized to 224 × 448 pixels (step S22).

[0036] Thereafter, these images were divided in half (step S23) to obtain a 224×224 pixel image D21 of the left half (right brain) and a 224×224 pixel image D22 of the right half (left brain) (step S24). The 224 x 224 pixel image D21 of the left half (right brain) was inverted to obtain ten 224 x 224 pixel inverted left half (right brain) images D23 (step S25). The inversion process of the 224 x 224 pixel image D21 of the left half (right brain), which is one of the halved images, in step S25 was performed to increase the data homogeneity with the right half (left brain) image D22. The images of multiple patients were used as input data, and the SBR values ​​of the cerebral dopamine transporter scintigraphy test corresponding to each patient were used as training data to train the first neural network NN1 to the fourth neural network NN4.

[0037] The first neural network NN1 is a neural network that uses the left half (right brain) image to predict and analyze the SBR value of the right brain. The second neural network NN2 is a neural network that uses the right half (left brain) image to predict and analyze the SBR value of the left brain. The third neural network NN3 is a neural network that uses the left half (right brain) image and the right half (left brain) image to predict and analyze the SBR values ​​of the right brain and the left brain, respectively. The fourth neural network NN4 is a neural network that uses the left half (right brain) image and the right half (left brain) image that are flipped left and right, and predicts and analyzes the SBR values ​​of the right brain and the left brain, respectively.

[0038] In this case, the signal intensity range of all DICOM images contained in each sample was different. Therefore, as preprocessing before inputting the data into the first neural network NN1 to the fourth neural network NN4, an additional normalization process was performed on the signal intensities between samples (step S26).

[0039] FIG. 7 is a diagram illustrating the normalization process of signal intensity between samples. Figure 7(a) shows the average signal intensity of each sample before normalization, (b) shows the maximum signal intensity of each sample before normalization, (c) shows the average signal intensity of each sample after normalization, and (d) shows the average and maximum signal intensities of each sample after normalization.

[0040] FIG. 8 is a table showing the mean error and mean square error for each of the three models created for each of the first neural network NN1 to the fourth neural network NN4. For each of the first neural network NN1 to the fourth neural network NN4, three models were created, one with and one without normalization of signal intensity between samples as additional preprocessing.

[0041] FIG. 9 is an explanatory diagram of a neural network used in the embodiment of the present invention. 3D Convolution Layer refers to a three-dimensional convolution layer. 3D Batch Normalization Layer refers to a three-dimensional normalization processing layer. The ReLU layer represents a normalized linear function and is a type of activation function. Flatten F1 converts the data into a vector in [1,K].

[0042] The 3D ConvBNActiBlock 50 is a layer block that combines a 3D Convolutional Layer 51, a 3D Batch Normalization 52, and a ReLU Layer 53 in this order. The 3D ConvActiBlock 60 is a layer block that combines a 3D Convolutional Layer 61 and a ReLU Layer 62 in this order.

[0043] For each of the first neural network NN1 to the fourth neural network NN4, neural networks were generated using the middle 10 DICOM images of the left half (right brain) and right half (left brain) of the sample, resized to 224 x 224 pixels, for multiple patients as input data, and the SBR values ​​of the right brain and left brain, respectively, as output data.

[0044] First, ten preprocessed DICOM images of size 224x224 are used as input data, and the first 3D ConvBNActiBlock 50a (kernel size is 3x3x3 (depth x height x width), stride is 1x2x2 (depth x height x width), padding is 1x1x1 (depth x height x width)) converts them from 1x10x224x224 to 64x10x112x112. The second 3D ConvBNActiBlock 50b (kernel size is 3x3x3, stride is 2x2x2, no padding) converts from 64x10x112x112 to 128x4x55x55. The third 3D ConvBNActiBlock 50c (kernel size is 3x3x3, stride is 3x3x3, no padding) converts from 128x4x55x55 to 512x1x18x18. The fourth 3D ConvBNActiBlock 50d (kernel size is 1x3x3, stride is 3x3x3, no padding) converts from 512x1x18x18 to 1024x1x6x6.

[0045] Next, convert from 1024×1×6×6 to 1024×1×1×1 in one 3DConvActiBlock 60 (kernel size is 1×6×6, stride is 1×1×1, no padding). Use flatten F1 to convert it into a 1x1024 vector. The 1024x512 fully connected layer 70a converts from 1x1024 to 1x512. The 512x128 fully connected layer 70b converts from 1x512 to 1x128. Finally, a 128x1 fully connected layer 70c outputs the SBR value as a 1x1 scalar.

[0046] In generating the neural network of this embodiment, Adam was used as the optimizer, and the weight of the L2 regularization term was 0.3. During training, the learning rate was set to 0.7 times the previous learning rate every 15 epochs. The loss function used was the mean squared error (MSE) loss function. The number of data samples was set to 284 for training data samples and 32 for validation (test) data samples for the first neural network 1NN and the second neural network NN2, which predict and analyze only one of the SBR values. The third neural network NN3 and the fourth neural network NN4, which predict and analyze both SBR values, had 567 training data samples and 63 validation (test) data samples. Each sample contains 10 intermediate DICOM images. The batch size for training and validation (test) data is 2.

[0047] The learning rates were 0.0008, 0.002, 0.003, 0.00001, 0.00008, 0.0001, 0.0002, 0.0003, 0.0004, 0.0005, 0.0006, 0.0007, 0.0009, 0.001, 0.005, 0.008, and 0.01. The development environment was Windows (registered trademark) 10, and the framework used was PyTorch.

[0048] FIG. 10 is a diagram illustrating the results of the SBR prediction analysis of the right brain when no normalization process is performed in the first neural network NN1. Figure 10(a) shows the distribution of predicted values ​​and true values ​​by the first neural network NN1, (b) shows the number of samples of the error between the predicted value and the true value, and (c) shows the distribution of the error by age. The normal value of SBR value may differ by age, so we also looked at the variation by age. The average error was 1.079211 and the mean squared error was 2.002554.

[0049] FIG. 11 is a diagram illustrating the SBR prediction analysis of the right brain when normalization processing is performed in the first neural network NN1. Figure 11(a) shows the distribution of predicted values ​​and true values ​​by the first neural network NN1, (b) shows the number of samples of error between predicted values ​​and true values, and (c) shows the distribution of error by age. The average error was 1.033450 and the mean squared error was 2.359399.

[0050] FIG. 12 is a diagram illustrating the results of the SBR prediction analysis of the left brain when no normalization process is performed in the second neural network NN2. Figure 12(a) shows the distribution of predicted values ​​and true values ​​by the second neural network NN2, (b) shows the number of samples of error between predicted values ​​and true values, and (c) shows the distribution of error by age. The average error was 0.958857 and the mean squared error was 1.446806.

[0051] FIG. 13 is a diagram illustrating the results of the SBR prediction analysis of the left brain when normalization processing is performed in the second neural network NN2. Figure 13(a) shows the distribution of predicted values ​​and true values ​​by the second neural network NN2, (b) shows the number of samples of error between predicted values ​​and true values, and (c) shows the distribution of error by age. The average error was 0.925819 and the mean squared error was 1.501469.

[0052] FIG. 14 is a diagram illustrating the results of SBR prediction analysis of the right and left brain when no normalization processing is performed in the third neural network NN3. Figure 14(a) shows the distribution of predicted values ​​and true values ​​by the third neural network NN3, (b) shows the number of samples of error between predicted values ​​and true values, and (c) shows the distribution of error by age. The average error was 1.495315 and the mean square error was 3.332895.

[0053] FIG. 15 is a diagram illustrating the results of the SBR prediction analysis of the right brain and the left brain when normalization processing is performed in the third neural network NN3. Figure 15(a) shows the distribution of predicted values ​​and true values ​​by neural network NN3, (b) shows the number of samples of error between predicted values ​​and true values, and (c) shows the distribution of error by age. The average error was 1.635853 and the mean squared error was 3.769612.

[0054] FIG. 16 is a diagram illustrating the results of SBR prediction analysis of the right brain and the left brain when no normalization processing is performed in the fourth neural network NN4. Figure 16(a) shows the distribution of predicted values ​​and true values ​​by the fourth neural network NN4, (b) shows the number of samples of error between predicted values ​​and true values, and (c) shows the distribution of error by age. The average error was 1.273068 and the mean square error was 2.794731.

[0055] FIG. 17 is a diagram illustrating the results of the SBR prediction analysis of the right brain and the left brain when normalization processing is performed in the fourth neural network NN4. Figure 17(a) shows the distribution of predicted values ​​and true values ​​by the fourth neural network NN4, (b) shows the number of samples of error between predicted values ​​and true values, and (c) shows the distribution of error by age. The average error was 1.441400 and the mean squared error was 3.066445.

[0056] As shown in FIGS. 10 to 17, it was confirmed that although the errors between the predicted and true SBR values ​​varied depending on age, most of the errors were small.

[0057] Figure 18 shows an example screen of software that predicts and analyzes the SBR value of a brain dopamine transporter scintigraphy test from image data of a head MRI test. As shown in Figure 18, first press the Import File button in the upper left corner and select the folder containing the samples, or drag and drop the folder into the left frame.

[0058] FIG. 19 is an example of a display of all DICOM images contained in the folder. As shown in Figure 19, by pressing the analysis button at the bottom right, a predictive analysis of the SBR value of the brain dopamine transporter scintigraphy test is started from the image data of the head MRI test.

[0059] FIG. 20 is a diagram showing an example of a display of the predicted SBR value displayed on the display unit of the diagnostic processing device in the embodiment of the present invention. As shown in Figure 20, the predicted analyzed SBR(R) and SBR(L) values ​​are displayed in the figure and also as numerical values ​​in the box below. The lower SBR(R) values ​​are indicated by dark circles, and the upper SBR(L) values ​​are indicated by light circles. The SBR value also decreases with normal aging. In FIG. 20, the central solid line represents the median, the upper dashed line represents the upper limit (median + 2 × standard deviation), and the lower dashed line represents the lower limit (median − 2 × standard deviation). Therefore, the area between the upper and lower dashed lines is the normal range, and this predictively analyzed SBR(R) and SBR(L) values ​​both show a decrease. The normal range for SBR decreases with age.

[0060] As described above, the performance evaluation results of the trained model 11 that was actually created clearly show that SBR(R) and SBR(L) values ​​can be predicted and analyzed from image data from a patient's head MRI examination, that the error between the predicted value and the true value varies depending on age, but is often small, and that the function of a diagnostic processing device is implemented by comparing the predicted value with the normal range, and that correlations exist between the multiple types of data included in the training data.

[0061] Although each embodiment has been described in detail above, it is not limited to a specific embodiment, and various modifications and changes are possible within the scope of the claims. It is also possible to combine all or some of the components of the above-described embodiments.

[0062] The diagnostic processing device 100 functions as a method for generating a trained model 11, and the method for generating the trained model 11 includes a storage step of storing image data from head MRI examinations of multiple patients and SBR values ​​from cerebral dopamine transporter scintigraphy examinations corresponding to each patient in a storage unit 12, and a step of generating a trained model 11 by deep learning using the image data from the head MRI examinations of multiple patients' heads stored in the storage step as input data and the SBR values ​​from the cerebral dopamine transporter scintigraphy examinations corresponding to each patient as training data. In generating the trained model 11, the image data from the head MRI examinations of multiple patients, which is the input data, and metadata such as data acquisition conditions for the cerebral dopamine transporter scintigraphy examinations corresponding to each patient, which is training data, may also be part of the input data.

[0063] The diagnostic processing device 100 has a computer program that is executed on the diagnostic processing device 100, and this computer program causes the diagnostic processing device 100 to execute a predictive analysis step of predicting and analyzing the SBR value of a brain dopamine transporter scintigraphy test of a patient using image data of the head MRI test of the patient as input data and a trained model 11 generated by deep learning using the SBR value of the brain dopamine transporter scintigraphy test corresponding to each patient as training data, and a display step of displaying the normal range of SBR and the predicted and analyzed SBR value.

[0064] In addition, the diagnostic processing method performed by the diagnostic processing device 100 includes a predictive analysis step in which image data from head MRI examinations of multiple patients is used as input data, and the SBR values ​​of the brain dopamine transporter scintigraphy examinations corresponding to each patient are used as training data to predict and analyze the SBR values ​​of the brain dopamine transporter scintigraphy examinations of the patients using a trained model 11 generated by deep learning, with the image data from the head MRI examinations of the patients as input, and a display step in which the normal range of SBR and the predicted and analyzed SBR values ​​are displayed.

[0065] The trained model 11 uses right brain images to predict and analyze the SBR value of the right brain using a first neural network NN1, and uses left brain images to predict and analyze the SBR value of the left brain using a second neural network NN2.

[0066] The trained model 11 uses right brain images and left brain images to predict and analyze the SBR values ​​of the right brain and the left brain using a third neural network NN3.

[0067] The trained model 11 uses a right-brain image or a left-brain image that has been flipped left and right, and predicts and analyzes the SBR values ​​of the right brain and the left brain using a fourth neural network NN4.

[0068] Computer programs and software that run on the diagnostic processing device 100 are read from the storage unit and executed by the CPU, thereby realizing various functions. [Explanation of symbols]

[0069] 100 Diagnostic processing device 10 Learning Department 11 Pre-trained models 12 Storage section 13 Input section 14 Predictive Analysis Department 15 Display section 16 Memory section NN1 First Neural Network NN2 Second Neural Network NN3 Third Neural Network NN4 Fourth Neural Network

Claims

1. a storage unit for storing image data of head MRI examinations of a plurality of patients' heads and SBR (Specific Binding Ratio) values ​​of brain dopamine transporter scintigraphy examinations corresponding to each patient; a learning unit that generates a trained model using deep learning with image data of head MRI examinations of multiple patients stored in the storage unit as input data and SBR values ​​of cerebral dopamine transporter scintigraphy examinations corresponding to each patient stored in the storage unit as training data, The trained model includes a fourth neural network that uses a right brain image and a left brain image to predict and analyze an SBR value of the right brain and an SBR value of the left brain, One of the right brain image and the left brain image is mirror-flipped, and the other image is not mirror-flipped. A learning device characterized by:

2. a storage step of storing image data of head MRI examinations of a plurality of patients and SBR values ​​of brain dopamine transporter scintigraphy examinations corresponding to each patient; and generating a trained model using deep learning with the image data of the head MRI examinations of the plurality of patients stored in the storage step as input data and the SBR values ​​of the brain dopamine transporter scintigraphy examinations corresponding to the respective patients stored in the storage step as training data, The image data includes one of the right brain image and the left brain image that is flipped left and right, and the other image that is not flipped left and right. A method for generating a trained model.

3. an input unit for inputting image data of a head MRI examination of a patient's head; a prediction analysis unit that uses image data from head MRI examinations of a plurality of patients as input data, and predicts the SBR value of the brain dopamine transporter scintigraphy examination of the patient based on the image data of the head MRI examination of the patient input from the input unit, using a trained model that has been deep-trained using the SBR value of the brain dopamine transporter scintigraphy examination corresponding to each patient as training data, The trained model includes a fourth neural network that uses a right brain image and a left brain image to predict and analyze an SBR value of the right brain and an SBR value of the left brain, One of the right brain image and the left brain image is mirror-flipped, and the other image is not mirror-flipped. A diagnostic processing device characterized by:

4. A display unit that displays the normal range of the SBR value and the SBR value predicted and analyzed by the prediction analysis unit.

4. The diagnostic processing device according to claim 3.

5. A computer program to be executed by a diagnostic processing device, The method causes the diagnostic processing device to execute a predictive analysis step of predicting and analyzing the SBR values ​​of the brain dopamine transporter scintigraphy tests of input patients based on the image data of the head MRI tests of the input patients, using a trained model generated by deep learning using image data of the head MRI tests of the input patients and SBR values ​​of the brain dopamine transporter scintigraphy tests corresponding to each patient as training data, The trained model includes a fourth neural network that uses a right brain image and a left brain image to predict and analyze an SBR value of the right brain and an SBR value of the left brain, One of the right brain image and the left brain image is mirror-flipped, and the other image is not mirror-flipped. A computer program characterized by:

6. A diagnostic processing method performed by a diagnostic processing device, an input step of inputting image data of a head MRI examination of a patient's head; and a predictive analysis step of predicting and analyzing the SBR values ​​of the brain dopamine transporter scintigraphy tests of the input patients based on the image data of the head MRI tests of the patients input in the input step, using a trained model generated by deep learning using image data of head MRI tests of the heads of multiple patients as input data and SBR values ​​of the brain dopamine transporter scintigraphy tests corresponding to each patient as training data, The trained model includes a fourth neural network that uses a right brain image and a left brain image to predict and analyze an SBR value of the right brain and an SBR value of the left brain, One of the right brain image and the left brain image is mirror-flipped, and the other image is not mirror-flipped. A diagnostic processing method comprising:

Images