Medical image diagnosis support device, medical image diagnosis support method and program

Abstract

To provide a medical image diagnosis support device capable of determining an extent of a disease while reducing a burden on a patient.SOLUTION: A pretreatment part 22 resizes a three-dimensional brain image so that the size of the three-dimensional brain image becomes a threshold. A model construction part 23 executes machine learning using the resized three-dimensional brain image for learning as learning data, and constructs an identification model for determining disease probability which is the probability of being affected by an object disease from the three-dimensional brain image. A determination part 24 determines disease probability of an inference object person by inputting an inference object image, which is the resized three-dimensional brain image showing the brain of the inference object person, to a learned model.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present disclosure relates to a medical image diagnosis support device, a medical image diagnosis support method, and a program. [Background technology]

[0002] Unlike many other physical illnesses, mental illnesses lack objective indicators that can be shared between therapists and patients. This makes it difficult for therapists to objectively explain to patients whether they have the illness, how to determine treatment, whether the illness has improved, and the duration of treatment, which can lead to problems such as delays or interruptions in treatment.

[0003] In recent years, advances in medical equipment such as magnetic resonance imaging (MRI) and computed tomography (CT) have made it possible to perform diagnostic imaging using high-resolution medical images. For example, it has been suggested that patients with schizophrenia, a type of psychiatric disorder, have reduced volumes of the bilateral hippocampus, amygdala, thalamus, and nucleus accumbens, as well as intracranial volume, compared to healthy individuals. Similarly, patients with Alzheimer's disease, another psychiatric disorder, are known to exhibit significant atrophy of the hippocampal region. Therefore, a voxel-based specific regional analysis system for Alzheimer's disease (VSRAD) has been developed as an early diagnostic aid for Alzheimer's disease by analyzing brain MRI images.

[0004] An MRI image is three-dimensional image data obtained by three-dimensionally superimposing multiple two-dimensional image data obtained by capturing images of a target such as the brain for each slice. Patent Document 1 discloses a diagnostic technique using such three-dimensional image data, which extracts a potential bleeding site from the three-dimensional image data, analyzes the bleeding site, and determines whether bleeding has occurred. Non-Patent Document 1 also discloses a technique for identifying brain diseases by individually processing slice images, which are two-dimensional image data for each slice of the three-dimensional brain image data. This technique extracts features, such as cortical thickness, that are medically believed to be correlated with disease from each slice image, and classifies the features using a linear classifier to determine whether a disease has occurred. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Japanese Patent Application Publication No. 2019-213785 [Non-patent literature]

[0006] [Non-Patent Document 1] W. Yassin, H. Nakatani, Y. Zhu, M. Kojima, K. Owada, H. Kuwabara, W. Gonoi, Y. Aoki, H. Takao, T. Natsubori, N. Iwashiro, K. Kasai, Y. Kano, O. Abe, H. Yamasue, and S. Koike, “Machine-learning classification using neuroimaging data in schizophrenia, autism, ultra-high risk and first-episode psychosis, “Translational Psychiatry, vol.10, no.278, pp.1-11, 2020. Summary of the Invention [Problem to be solved by the invention]

[0007] Since it is difficult to clearly distinguish whether a mental illness such as schizophrenia is present or absent, it is important to determine the severity of the illness. However, the technology described in Patent Document 1 only determines whether bleeding is present or absent, making it difficult to apply to mental illnesses. Similarly, the technology described in Non-Patent Document 1 also only determines whether a disease is present or absent using a linear classifier, making it difficult to apply to mental illnesses, as with the technology described in Patent Document 1.

[0008] Furthermore, in imaging diagnosis using MRI images, high-resolution image data is required to accurately diagnose diseases, but the higher the resolution, the longer the imaging time increases, which increases the burden on patients.

[0009] The present disclosure has been made in consideration of the above-mentioned problems, and aims to provide a medical image diagnosis support device, a medical image diagnosis support method, and a program that are capable of determining the extent of disease while reducing the burden on patients. [Means for solving the problem]

[0010] A medical image diagnosis support device according to one aspect of the present disclosure is a medical image diagnosis support device that supports the diagnosis of a predetermined target disease related to the brain using a three-dimensional brain image showing the brain, and includes: a preprocessing unit that generates a processed image by resizing the three-dimensional brain image so that the size of the three-dimensional brain image becomes a threshold value; a model construction unit that performs machine learning using a plurality of the processed images showing the brain of a training subject as training data and constructs a trained model that determines the disease probability, which is the probability that the subject has the target disease, from the processed images; and a determination unit that inputs an inference target image, which is the processed image showing the brain of the inference subject, into the trained model and determines the disease probability of the inference subject. [Effects of the Invention]

[0011] According to the present invention, it is possible to accurately determine the extent of a disease while reducing the burden on the patient. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 1 illustrates a diagnosis support system according to an embodiment of the present disclosure. [Figure 2] 1 is a diagram illustrating an example of the configuration of a medical image processing apparatus according to an embodiment of the present invention. [Figure 3] FIG. 1 is a diagram showing an example of a three-dimensional brain image. [Figure 4] FIG. 10 is a diagram showing another example of a three-dimensional brain image. [Figure 5] FIG. 10 is a diagram showing an example of a correspondence relationship between a patient and a disease probability. [Figure 6] FIG. 10 is a diagram showing an example of a correspondence relationship between healthy subjects and disease probabilities. [Figure 7] FIG. 10 is a diagram illustrating an example of display information. [Figure 8] FIG. 10 is a diagram showing another example of display information. [Figure 9] 10 is a flowchart illustrating an example of processing by a medical image processing apparatus. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings. [Example]

[0014] Fig. 1 is a diagram illustrating a diagnosis support system according to an embodiment of the present disclosure. The diagnosis support system 100 illustrated in Fig. 1 includes a medical image processing device 1, a three-dimensional image capturing device 2, and an image storage server 3. The medical image processing device 1, the three-dimensional image capturing device 2, and the image storage server 3 are communicably connected to each other via a network 4. Note that the diagnosis support system 100 illustrated in Fig. 1 is merely an example, and, for example, some or all of the devices 1 to 3 may be integrated. Furthermore, the network 4 may be wired or wireless, and the devices 1 to 3 may be directly connected without going through the network 4.

[0015] The medical image processing device 1 is a medical image diagnosis support device that analyzes medical images of a diagnostic target region of a subject to be diagnosed and supports the determination of a predetermined target disease related to the diagnostic target region. In this embodiment, the diagnostic target region is the brain, and the target disease is a mental disorder such as schizophrenia, Alzheimer's disease, developmental disorder, depression, bipolar disorder, dementia, or intellectual disability. However, the diagnostic target region is not limited to the brain, and the target disease is not limited to a mental disorder.

[0016] The medical image processing device 1 is realized, for example, by installing and executing a medical image processing program on a computer. The computer may be a workstation or personal computer operated directly by a user, such as a doctor making a diagnosis, or a server computer connected to the workstation or personal computer via a network. The medical image processing program may be recorded on a recording medium such as a DVD (Digital Versatile Disc) or CD-ROM (Compact Disc Read Only Memory) and distributed, and then installed on a computer from the recording medium. Alternatively, the medical image processing program may be stored in a storage device of a server computer connected to a network or in network storage in an externally accessible state, and downloaded and installed on a user's computer upon request.

[0017] The three-dimensional imaging device 2 is a device that captures an image of a diagnostic target region to generate a medical image showing the diagnostic target region. The three-dimensional imaging device 2 is, for example, an MRI device, a CT device, or a PET (Positron Emission Tomography) device, and the medical image is an image of the diagnostic target region. In this embodiment, the three-dimensional imaging device 2 is an MRI device. There may be multiple types of medical images. In this embodiment, the medical images include T1-weighted images and VSRAD images, which are images captured for VSRAD. T1-weighted images have the advantage of easily capturing anatomical structures and easily detecting morphological abnormalities. The medical images are not limited to T1-weighted images and VSRAD images, and may also be, for example, T2-weighted images, Fluid Attenuated Inversion Recovery (FLAIR) images, Diffusion Weighted Images (DWI) images, Magnetic Resonance Angiography (MRA) images, Maximum Intensity Projection (MIP) images, functional MRI (f-MRI) images, etc.

[0018] The image storage server 3 is a device that stores medical images generated by the three-dimensional imaging device 2. Specifically, the image storage server 3 is a computer device that stores and manages various types of data, and is equipped with a large-capacity external storage device (not shown). The image storage server 3 uses database management software to acquire medical images generated by the three-dimensional imaging device 2 via the network 4, and stores and manages the images in the large-capacity external storage device. The storage format of the medical images and the communication format between the devices 1 to 3 are based on protocols such as DICOM (Digital Imaging and Communication in Medicine).

[0019] Fig. 2 is a diagram showing an example of the configuration of a medical image processing device 1. The medical image processing device 1 shown in Fig. 1 has a memory 11, a storage 12, and a CPU (Central Processing Unit) 13 as a standard workstation configuration. The medical image processing device 1 is also connected to an input unit 14 and an output unit 15. The input unit 14 includes a keyboard, a mouse, etc., and receives various information from a user. The output unit 15 includes a display unit such as a display, and outputs (for example, displays) various information.

[0020] The memory 11 stores a medical image processing program 20, which is a computer program that defines the operation of the CPU 13. The storage 12 includes a hard disk drive and the like. The storage 12 stores medical images acquired from the image storage server 3 via the network 4, as well as various information used and generated by the CPU 13.

[0021] The CPU 13 is a processor that reads the medical image processing program 20 stored in the memory 11 and executes the read medical image processing program 20 to realize an image acquisition unit 21, a preprocessing unit 22, a model construction unit 23, a judgment unit 24, an identification basis visualization unit 25, and a display control unit 26.

[0022] The processor is not limited to a CPU, but may be a programmable logic device (PLD), such as a field programmable gate array (FPGA), whose circuit configuration can be changed after manufacture. Alternatively, a dedicated electrical circuit having a circuit configuration specifically designed for executing a specific process, such as an application specific integrated circuit (ASIC), may be used as the processor.

[0023] The image acquisition unit 21 acquires medical images from the image storage server 3. If the medical images are stored in the storage 12, the image acquisition unit 21 may acquire the medical images from the storage 12. Medical images include training medical images used as learning data for machine learning, and inference medical images used to make inferences about a target disease using a trained model constructed by machine learning. Training medical images are images of the brain of a training subject, who is a subject whose diagnosis has been completed, and are labeled with a correct answer indicating whether or not the subject is suffering from the target disease. Inference medical images are images of the brain of an inference subject whose target disease is to be determined.

[0024] As preprocessing, the preprocessing unit 22 generates a 3D brain image, which is a processed image obtained by performing various image processing on the medical image. Specifically, the preprocessing unit 22 converts the medical image into a 3D brain image composed of multiple slice images arranged in a predetermined direction. As a result, the 3D brain image becomes an image in which pixels are arranged in three different directions. In addition, in the process of converting the medical image into the 3D brain image, the preprocessing unit 22 performs image processing such as normalization, resizing, and data expansion of the medical image. Note that at least a part of the preprocessing performed by the preprocessing unit 22 may be performed by another device, such as the 3D image capturing device 2.

[0025] Normalization is a process of converting the voxel values ​​of a three-dimensional brain image according to a predetermined rule. In this embodiment, the pre-processing unit 22 performs normalization by converting the voxel values ​​of the three-dimensional brain image so that the average of the voxel values ​​of the three-dimensional brain image is 0 and the variance is 1. However, the pre-processing unit 22 may also use a process based on the Min-Max method for normalizing the voxel values ​​of the three-dimensional brain image so that the minimum value of the voxel value of the three-dimensional brain image is 0 and the maximum value is 1.

[0026] Resizing is a process of converting the size of a 3D brain image. In this embodiment, the sizes are the sizes in three different directions (X-axis, Y-axis, and Z-axis), and resizing is performed so that the size in each direction matches a threshold value. For example, if the size in each direction is smaller than the threshold value, the preprocessing unit 22 performs upsampling using numerical interpolation in that direction. If the size in each direction is larger than the threshold value, the preprocessing unit 22 performs downsampling using numerical interpolation in that direction to convert the size in each direction to the threshold value. The threshold values ​​corresponding to the sizes in each direction may be different from each other. For example, because reducing the imaging time in the Z-axis direction of the subject is more effective in reducing the burden on the patient, the threshold values ​​corresponding to the X-axis and Y-axis directions may be larger than the threshold value corresponding to the Z-axis direction.

[0027] In this embodiment, the numerical interpolation used for resizing is nonlinear interpolation, which is numerical interpolation using nonlinear transformation. Nonlinear interpolation is, for example, spline interpolation, which uses a spline function to nonlinearly interpolate image positions. Nonlinear interpolation makes it possible to perform interpolation while maintaining spatial smoothness.

[0028] Data augmentation is a process for increasing the number of medical images by generating new medical images by performing image conversion on the medical images. Examples of image conversion processes include rotation, shift, scaling, and brightness conversion. Rotation is a process for rotating an image around at least one of the X-axis, Y-axis, and Z-axis directions as the rotation axis. Shifting is a process for translating an image along at least one of the X-axis, Y-axis, and Z-axis directions. Scaling is a process for scaling an image similarly along at least one of the X-axis, Y-axis, and Z-axis directions. Brightness conversion is a process for adjusting the brightness of a medical image by smoothing using a Gaussian filter or the like, and sharpening by increasing edge weights. Data augmentation is not limited to these examples, and may also include horizontal flipping or random cropping as image conversion processes.

[0029] The preprocessing unit 22 may also align the 3D brain images as preprocessing. Alignment is a general term for processing that addresses the problem of brain tissue positions differing from image to image due to differences in imaging environments and individual differences in brain shape, and is performed using rotation, shift, scaling, and other processes. Alignment includes standardization of brain morphology in the 3D brain images to a standard brain morphology using SPM12 (Statistical Parametric Mapping 12) and other methods.

[0030] Specifically, the registration can be performed using AC-PC (Anterior Commissure-Posterior Commissure) registration, realignment processing, standardization processing, etc. AC-PC registration is a process that sets the origin on the brain's AC (anterior commissure) and aligns the 3D brain image so that it is parallel to the AC-PC line to prevent incorrect processing due to misalignment of the origin of each 3D brain image. Realignment is a process that corrects misalignment of medical images due to head movement caused by the subject's body movement and heartbeat during imaging. It corrects the position of the head in the 3D brain image by performing rigid body transformation according to the amount of translation and rotation calculated using the least squares method, etc. Standardization is a process that aligns a standard brain image with another 3D brain image. A standard brain image is a brain image obtained by averaging multiple brain images in which brain regions are extracted from multiple MRI images acquired by an MRI device to image the heads of multiple healthy subjects. The standard brain image may be an image created using computer graphics or may be an image obtained by imaging the brain of a healthy subject. The standard brain image is stored in the image storage server 3 or storage 12, and the preprocessing unit 22 obtains the standard brain image from the image storage server 3 or storage 12.

[0031] In this embodiment, the preprocessing unit 22 performs registration on each of the plurality of data-extended three-dimensional brain images, but may perform data extension on the registered three-dimensional brain images.

[0032] 3 and 4 are diagrams showing examples of three-dimensional brain images converted by the preprocessing unit 22. Specifically, Fig. 3 shows an example of a three-dimensional brain image B0 converted from a T1-weighted image, and Fig. 4 shows an example of a three-dimensional brain image C0 converted from a VSRAD image.

[0033] As shown in Figures 3 and 4, the three-dimensional brain images B0 and C0 show a skull 31 and a brain parenchyma 32. The brain parenchyma 32 includes various anatomical regions such as the ventricles and cisterns. In Figures 3 and 4, the ventricles 33 are shown as anatomical regions, and other anatomical regions are omitted.

[0034] Returning to the explanation of Figure 2, the model construction unit 23 performs machine learning using the 3D brain images converted from the training medical images by the preprocessing unit 22 as training data, and constructs a discrimination model, which is a trained model that determines the disease probability, which is the probability of suffering from the target disease, from the 3D brain image. The model construction unit 23 may increase the number of training rounds by cross-validation in order to improve the generalization performance of the discrimination model and prevent over-learning. Furthermore, the discrimination accuracy (evaluation index) of the discrimination model may be, for example, AUC (Area Under the Curve), TPR (True Positive Rage), or FPR (False Positive Rate).

[0035] In this embodiment, the model construction unit 23 constructs a classification model using a three-dimensional classifier. Unlike a two-dimensional classifier that identifies image features for each slice image, a three-dimensional classifier is a classifier that can identify image (brain) features using a three-dimensional image as is, such as a three-dimensional convolutional neural network (3D-CNN). A three-dimensional classifier generally has higher computational costs than a two-dimensional classifier, but improves classification accuracy. In this embodiment, the three-dimensional classifier is a 3D-CNN. A 3D-CNN is a nonlinear classifier that analyzes three-dimensional brain images down to the three-dimensional image level, and can calculate and output a disease probability as a continuous value. The disease probability serves as an objective index that supports the diagnosis of the target disease.

[0036] The determination unit 24 inputs an inference target image, which is a 3D brain image converted from the medical image for inference by the preprocessing unit 22, into the discrimination model constructed by the model construction unit 23, and calculates and outputs the disease probability of the target disease as a continuous value. At this time, the determination unit 24 may determine whether the inference target, which is the subject, is a healthy individual based on the disease probability. For example, when a sigmoid function is used in the output layer of the 3D-CNN, the determination unit 24 determines that the inference target is a patient suffering from the target disease if the disease probability is 0.5 or more and 1.0 or less, and determines that the subject is a healthy individual if the disease probability is 0 or more and less than 0.5. The output layer of the 3D-CNN is not limited to a sigmoid function, and other functions such as a softmax function may also be used.

[0037] Fig. 5 is a diagram showing the relationship between diseased individuals and disease probability calculated from 3D brain images showing the brains of the diseased individuals, and Fig. 6 is a diagram showing the relationship between healthy individuals and disease probability calculated from 3D brain images showing the brains of healthy individuals. In Figs. 5 and 6, the horizontal axis represents disease probability, and the vertical axis in Fig. 5 represents the number of subjects (the number of diseased individuals or the number of healthy individuals). As shown in Figs. 5 and 6, disease probability is expressed as a continuous value rather than a discrete value such as 0 or 1, and therefore serves as an objective index useful for diagnosing mental illness.

[0038] Returning to the explanation of Figure 2, the discrimination basis visualization unit 25 is a visualization unit that identifies a discrimination basis region in the 3D brain image, which is a region used by the determination unit 24 to calculate the disease probability. Specifically, the discrimination basis visualization unit 25 identifies the discrimination basis region using 3D-Grad-CAM (Three-Dimensional Gradient-weighted Class Activation Mapping), which is a type of XAI (eXplainable Artificial Intelligence), a technology that analyzes the determination basis that forms the basis for determination by a machine learning model. 3D-Grad-CAM is a technology that has high class discrimination performance and provides explainable grounds for the results output by a machine learning model, etc.

[0039] The display control unit 26 generates display information indicating at least one of the disease probability calculated by the determination unit 24 and the identification basis region specified by the identification basis visualization unit 25, and displays it on the output unit 15.

[0040] 7 and 8 are diagrams showing examples of display information. Specifically, Fig. 7 is an example of display information showing an identification basis region in a three-dimensional brain image B0, and an example of display information showing an identification basis region in a three-dimensional brain image C0.

[0041] The display information 40 shown in Fig. 7 includes a slice brain image 41 including an identification basis region from the 3D brain image B0, specific information 42 for identifying the slice brain image, and explanatory information 43 explaining the identification basis region. The slice brain image 41 shows the skull 31 and brain parenchyma 32, and further shows an identification basis region 34 within the brain parenchyma 32. The specific information 42 is text information indicating the number of the displayed slice brain image 41 among all slice brain images. The explanatory information 43 is text information explaining the region corresponding to the identification basis region 34. Note that the specific information 42 and explanatory information 43 are not limited to text information and may be information such as a symbol or an icon.

[0042] 8 includes a slice brain image 51, specific information 52, and explanatory information 53, similar to the display information 40. The slice brain image 51, specific information 52, and explanatory information 53 correspond to the slice brain image 41, specific information 42, and explanatory information 43 shown in FIG.

[0043] The display information is not limited to showing one slice brain image as shown in Figures 7 and 8, but may show multiple slice brain images, or may show the three-dimensional brain images B0 and C0 as they are.

[0044] FIG. 9 is a flowchart for explaining an example of processing by the medical image processing apparatus 1.

[0045] First, the learning phase F1 shown on the left side of Fig. 9 is executed. In the learning phase F1, the image acquisition unit 21 acquires learning medical images (T1 weighted images and VSRAD images) from the image storage server 3 (step ST1).

[0046] The preprocessing unit 22 converts each of the learning medical images acquired by the image acquisition unit 21 into three-dimensional images to create three-dimensional brain images B0 and C0 (step ST2).

[0047] The preprocessing unit 22 normalizes each of the three-dimensional brain images B0 and C0 so that the average of the voxel values ​​of each of the three-dimensional brain images B0 and C0 is 0 and the variance is 1 (step ST3). However, normalization (the processing of step ST3) does not have to be performed.

[0048] The preprocessing unit 22 resizes each of the three-dimensional brain images B0 and C0 so that the sizes of the three-dimensional brain images B0 and C0 in the X-axis direction, Y-axis direction, and Z-axis direction become equal to threshold values ​​(step ST4). Specifically, the preprocessing unit 22 increases the size to the threshold value by performing upsampling using numerical interpolation in directions where the size is less than the threshold value, and reduces the size to the threshold value by performing downsampling using numerical interpolation in directions where the size is greater than the threshold value. Note that the preprocessing unit 22 does not need to resize the image in a direction where the size is equal to the threshold value.

[0049] The preprocessing unit 22 performs image conversion processing on at least a part of each of the three-dimensional brain images B0 and C0, thereby increasing the number of the three-dimensional brain images B0 and C0 (step ST5). However, data expansion, which is the processing of step ST5, may not be performed.

[0050] The model construction unit 23 inputs the 3D brain images as learning data to the 3D-CNN, causes the 3D-CNN to perform machine learning, and constructs a discriminative model, which is a trained model that determines the disease probability of the target disease from the 3D brain images B0 and C0 (step ST6), thereby completing the learning phase F1. In this embodiment, the discriminative model determines a single disease probability from both the 3D brain images B0 and C0.

[0051] In the example shown in the figure, for simplicity, normalization, resizing, data expansion, etc. are performed on the three-dimensional brain images B0 and C0 generated by three-dimensionalization, but normalization, resizing, and data expansion may also be performed in the process of generating the three-dimensional brain images B0 and C0.

[0052] Next, inference phase F2 shown on the right side of Fig. 9 is executed. In inference phase F2, first, the image acquisition unit 21 acquires medical images for inference (T1-weighted images and VSRAD images) from the image storage server 3 (step ST11). Thereafter, the pre-processing unit 22 performs the same processes as steps ST2 to ST4 of learning phase F1 on the medical images for inference to generate three-dimensional brain images B0 and C0 for inference (steps ST12 to ST14).

[0053] Thereafter, the judgment unit 24 inputs the three-dimensional brain images B0 and C0 for inference into the discrimination model constructed by the model construction unit 23 in step ST6, and calculates the disease probability that the subject is suffering from the target disease (step ST15).

[0054] The discrimination basis visualizing unit 25 analyzes the process of calculating the disease probability using the discrimination model, and identifies discrimination basis regions that are regions used to calculate the disease probability in the three-dimensional brain images B0 and C0 (step ST16).

[0055] Then, the display control unit 26 generates display information indicating at least one of the disease probability calculated by the judgment unit 24 and the identification basis area identified by the identification basis visualization unit 25, and displays it on the output unit 15 (step ST17), thereby completing the inference phase F2.

[0056] As described above, according to this embodiment, the preprocessing unit 22 resizes the 3D brain image so that the sizes of the 3D brain image in the X-axis, Y-axis, and Z-axis directions meet threshold values. The model construction unit 23 performs machine learning using the resized 3D brain image for training as training data to construct a discrimination model that determines the disease probability, which is the probability that the subject has a target disease, from the 3D brain image. The determination unit 24 inputs the inference target image, which is a resized 3D brain image showing the brain of the subject to the trained model, to determine the disease probability of the subject. Therefore, the disease probability of the subject having the target disease is determined from the processed image, which has been resized so that the 3D brain image has an appropriate size. Therefore, even if a small image is acquired, an image of an appropriate size can be generated, thereby shortening the imaging time and reducing the burden on the patient. Furthermore, since the disease probability is determined rather than the presence or absence of the disease, the severity of the disease can be determined.

[0057] In this embodiment, the 3D brain image is resized using a numerical interpolation method, which allows for smooth interpolation and allows resizing while maintaining spatial smoothness.

[0058] Furthermore, in this embodiment, the three-dimensional brain image is resized using a numerical interpolation method with a nonlinear function, which makes it possible to perform smoother interpolation more appropriately.

[0059] In this embodiment, the preprocessing unit 22 resizes the three-dimensional brain image so that the sizes of the three-dimensional brain image in the X-axis direction, the Y-axis direction, and the Z-axis direction each reach a threshold value, thereby enabling the three-dimensional image to be resized to a more appropriate size.

[0060] In addition, in this embodiment, the threshold values ​​corresponding to the different directions are different from each other, which makes it possible to resize the three-dimensional image to a more appropriate size.

[0061] In this embodiment, the identification basis visualization unit 25 identifies the identification basis region used to calculate the disease probability in the 3D brain image, which allows doctors and others to determine the target disease by focusing on the identification basis region information, thereby supporting appropriate diagnosis of the target disease.

[0062] Furthermore, the model construction unit 23 performs machine learning using 3D-CNN, which makes it possible to determine the probability of disease while maintaining geometric information without converting the three-dimensional brain image into feature quantities.

[0063] It should be noted that the present disclosure is not limited to the above-described examples and includes various modifications. For example, the examples have been described in detail to clearly explain the present invention, and are not necessarily limited to those having all of the described configurations. For example, although the present examples are directed to medical images, the present disclosure is not limited to medical images and can also be applied to material diagnosis using images of materials.

[0064] Furthermore, some of the configurations of the embodiments may be added, deleted, or replaced. Furthermore, the above-described configurations, functions, processing units, processing groups, etc. may be implemented in hardware, in part or in whole, by designing them as integrated circuits, for example. Furthermore, the above-described configurations, functions, etc. may be implemented in software by a processor interpreting and executing a program that implements each function. Information such as programs, tables, and files that implement each function may be stored in a storage device such as a memory, hard disk, or solid-state drive (SSD), or in a recording medium such as an IC card, SD card, or DVD.

[0065] Furthermore, one processing unit may be configured with one of these various processors, or may be configured with a combination of two or more processors of the same or different types (for example, multiple FPGAs, or a combination of a CPU and an FPGA). Also, multiple processing units may be configured with one processor. As an example of configuring multiple processing units with one processor, first, there is a form in which one processor is configured with a combination of one or more CPUs and software, as typified by computers such as client and server, and this processor functions as multiple processing units. Second, there is a form in which a processor is used to realize the functions of an entire system including multiple processing units with a single IC (Integrated Circuit) chip, as typified by a system on chip (SoC). In this way, the various processing units are configured with one or more of the above various processors as a hardware structure.

[0066] Furthermore, the hardware structure of these various processors is, more specifically, an electronic circuit that combines circuit elements such as semiconductor elements. [Explanation of symbols]

[0067] 1: Medical image processing device 2: 3D image capturing device 3: Image storage server 21: Image acquisition unit 22: Preprocessing unit 23: Model construction unit 24: Determination unit 25: Identification basis visualization unit 26: Display control unit

Claims

1. A medical image diagnosis support device that supports the diagnosis of a predetermined target disease related to the brain using a three-dimensional brain image showing the brain, a pre-processing unit that generates a processed image by resizing the three-dimensional brain image so that the size of the three-dimensional brain image matches a threshold value; a model construction unit that performs machine learning using a plurality of the processed images showing the brain of a training subject as training data, and constructs a trained model that determines a disease probability, which is the probability that the subject has the target disease, from the processed images; and A determination unit that inputs an inference target image, which is the processed image showing the brain of the inference target, into the trained model and determines the disease probability of the inference target; A visualization unit that identifies an identification basis area, which is an area in the inference target image that is used to calculate the disease probability in the trained model, The medical image diagnosis support device, wherein the three-dimensional brain image includes a plurality of types of images.

2. the three-dimensional brain image is an image in which pixels are arranged in three different directions, The medical image diagnosis support device according to claim 1 , wherein the size is a size in each direction.

3. The medical image diagnosis support device according to claim 1 , wherein the preprocessing unit resizes the three-dimensional brain image using a numerical interpolation method.

4. The medical image diagnosis support device according to claim 2 , wherein the preprocessing unit resizes the three-dimensional brain image using a numerical interpolation method that uses a nonlinear function.

5. The medical image diagnosis support device according to claim 2 , wherein the preprocessing unit resizes the three-dimensional brain image so that the size of each direction of the three-dimensional brain image meets a threshold value.

6. The medical image diagnosis support device according to claim 5 , wherein the threshold values ​​corresponding to the respective directions are different from each other.

7. The medical image diagnosis support device according to claim 1 , wherein the model construction unit performs the machine learning using a 3D-CNN (Three-Dimensional-Convolutional Neural Network).

8. A medical image diagnosis support method using a medical image diagnosis support device that supports the diagnosis of a predetermined target disease related to the brain using a three-dimensional brain image showing the brain, comprising: generating a processed image by resizing the three-dimensional brain image so that the size of the three-dimensional brain image becomes a predetermined value; performing machine learning using a plurality of the processed images showing the brain of a training subject as training data, and constructing a trained model that determines a disease probability, which is the probability that the subject has the target disease, from the processed images; An inference target image, which is the processed image showing the brain of the inference target, is input into the trained model to determine the disease probability of the inference target; Identifying an identification basis region in the inference target image that is a region used in calculating the disease probability in the trained model; A medical image diagnosis support method, wherein the three-dimensional brain image includes multiple types of images.

9. A program for assisting in the diagnosis of a predetermined target disease related to the brain using a three-dimensional brain image showing the brain, generating a processed image by resizing the three-dimensional brain image so that the size of the three-dimensional brain image becomes a predetermined value; a step of performing machine learning using a plurality of the processed images showing the brain of a training subject as training data, and constructing a trained model that determines a disease probability, which is the probability that the subject has the target disease, from the processed images; A procedure of inputting an inference target image, which is the processed image showing the brain of an inference target, into the trained model and determining the disease probability of the inference target; and a step of identifying an identification basis region in the inference target image that is used in calculating the disease probability in the trained model, The program, wherein the three-dimensional brain image includes multiple types of images.

Images