Lesion detection method and program

Abstract

To detect a lesion region at constant accuracy regardless of an imaging position of a tomographic image, in a human body.SOLUTION: A lesion detection method is configured to: acquire a tomographic image 22a obtained by imaging a first imaging position of a human body 21; select any one of a machine learning model 2a and a machine learning model 2b identifying whether a region is a region of lesion, with an image block obtained by dividing an input tomographic image as a unit, on the basis of the number of the tomographic images imaged at a second imaging position corresponding to a first imaging position, in human bodies 11a, 11b, out of multiple tomographic images including a lesion region respectively, the tomographic images being used as learning data when generating the machine learning model 2a for identifying whether the region is a region of lesion, with a pixel included in the input tomographic image as a unit, and the tomographic images being obtained by imaging insides of the human bodies 11a, 11b; then, use the selected machine learning model to detect the lesion region from the tomographic image 22a.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a lesion detection method and a lesion detection program. [Background technology]

[0002] Medical images such as those obtained by CT (Computed Tomography) and MRI (Magnetic Resonance Imaging) are widely used to diagnose various diseases. When using medical images for image diagnosis, doctors must interpret a large number of images, which places a heavy burden on them. For this reason, there is a demand for technology that uses computers to assist doctors in some way in their diagnostic work.

[0003] An example of such a technology is a technology that detects lesion areas from medical images using a trained model (machine learning model) generated by machine learning. For example, a system for cardiac segmentation has been proposed that trains a neural network integrated with LSTM (Long-Short-Term Memory). [Prior art documents] [Patent documents]

[0004] [Patent Document 1] US Patent Application Publication No. 2019 / 0223725 Summary of the Invention [Problem to be solved by the invention]

[0005] One example of a machine learning model is one that identifies whether each pixel in a captured tomographic image is a region of a specific lesion. In a classification process using this machine learning model, when a single tomographic image is input to the machine learning model, the lesion region is identified pixel by pixel from the tomographic image.

[0006] The problem with such machine learning models is that the accuracy of classification may decrease depending on the position in the human body where the tomographic image is taken. For example, when generating such a machine learning model, tomographic images of a human body that have actually been captured are used as training data. Normally, tomographic images are captured from a certain range inside every human body at equal intervals, so the number of tomographic images used as training data is constant regardless of the position in the human body. On the other hand, the frequency of appearance of a lesion area may be biased depending on the position inside the human body. Therefore, among the tomographic images used as training data, the number of tomographic images that include a lesion area may be biased depending on the position inside the human body. For tomographic images captured at a position where a sufficient number of tomographic images are not used as training data, the accuracy of identifying lesions may be reduced.

[0007] In one aspect, the present invention aims to provide a lesion detection method and a lesion detection program capable of detecting a lesion area with a certain degree of accuracy regardless of the position at which a tomographic image is taken inside a human body. [Means for solving the problem]

[0008] One proposal provides a lesion detection method in which a computer executes the following process: In this lesion detection method, the computer acquires a first tomographic image captured at a first imaging position inside a first human body, selects either the first machine learning model or a second machine learning model that identifies whether or not an area of ​​a specific lesion is included in each image block obtained by dividing the input tomographic image into a certain size, based on the number of tomographic images captured at a second imaging position corresponding to the first imaging position inside one or more second human bodies among a plurality of tomographic images each including an area of ​​a specific lesion, which have been used as learning data when generating, by machine learning, a first machine learning model that identifies whether or not an area of ​​a specific lesion is included in each pixel included in the input tomographic image, and detects an area of ​​a specific lesion from the first tomographic image using the machine learning model selected from the first machine learning model and the second machine learning model.

[0009] Also, one proposal provides a lesion detection method in which a computer executes the following process. In this lesion detection method, the computer acquires a plurality of tomographic images capturing the inside of one or more second human bodies, and a plurality of correct answer data corresponding to each of the plurality of tomographic images and indicating whether or not each pixel on the corresponding tomographic image is a region of a specific lesion, classifies the plurality of tomographic images into one of a plurality of first divided regions depending on which divided region each of the plurality of tomographic images was captured in among a plurality of first divided regions obtained by dividing a predetermined organ region in one or more second human bodies into a predetermined number of divisions in a predetermined direction, inputs the plurality of tomographic images to a first machine learning model that identifies whether or not a region is a region of a specific lesion using pixels included in the input tomographic images as units, executes a discrimination process, and compares the discrimination result with the plurality of correct answer data, thereby classifying the plurality of tomographic images into one of a plurality of first divided regions depending on which divided region each of the plurality of tomographic images was captured in, and executes a discrimination process by inputting the plurality of tomographic images to a first machine learning model that identifies whether or not a region is a region of a specific lesion using pixels included in the input tomographic images as units ... the first tomographic image is acquired, and a third divided area in which the first tomographic image was acquired is identified from a plurality of second divided areas obtained by dividing an organ area in the first human body into a predetermined number of divisions in a predetermined direction; based on the index calculated for a fourth divided area corresponding to the third divided area from the plurality of first divided areas, either a first machine learning model or a second machine learning model that identifies whether or not an area of ​​a specific lesion is present in each image block obtained by dividing the input tomographic image into image blocks of a certain size is selected; and the first tomographic image is detected from the first tomographic image using the machine learning model selected from the first machine learning model and the second machine learning model.

[0010] Furthermore, in one aspect, a lesion detection program is provided that causes a computer to execute a process similar to the lesion detection method described above. Effect of the Invention

[0011] In one aspect, the lesion area can be detected with a certain degree of accuracy regardless of the position within the human body where the tomographic image is taken. [Brief description of the drawings]

[0012] [Figure 1] 1 is a diagram illustrating an example of the configuration and processing of a lesion detection device according to a first embodiment; [Diagram 2] FIG. 11 illustrates an example of a configuration of a diagnosis support system according to a second embodiment. [Diagram 3] FIG. 1 illustrates a first example of a trained model for lesion identification. [Figure 4] FIG. 13 illustrates a second example of a trained model for lesion identification. [Diagram 5] FIG. 13 is a diagram showing the bias in the number of pieces of learning data in the organ region. [Figure 6] FIG. 13 is a diagram showing an example of selection of a lesion identification model. [Figure 7] FIG. 13 is a diagram showing a specific example of model selection according to the number of learning data of a lesion area. [Figure 8] FIG. 2 is a diagram illustrating an example of the configuration of processing functions included in a diagnosis support device. [Figure 9] FIG. 4 is a diagram illustrating an example of data stored in a learning data storage unit. [Figure 10] FIG. 4 is a diagram illustrating an example of data stored in a setting data storage unit. [Figure 11] FIG. 11 is a diagram illustrating an example of data stored in a tabulated data storage unit. [Figure 12] FIG. 4 is a diagram illustrating an example of data stored in a model data storage unit. [Figure 13] 11 is an example of a flowchart illustrating a procedure for a data counting process. [Figure 14] 13 is an example of a flowchart illustrating a procedure for a model determination process. [Figure 15] 13 is an example of a flowchart illustrating a procedure for a model generation process. [Figure 16] 13 is an example of a flowchart showing a procedure of a lesion identification process. [Figure 17] FIG. 13 is a diagram illustrating an example of a configuration of processing functions included in a diagnosis support device according to a third embodiment. [Figure 18] FIG. 11 is a diagram illustrating an example of data stored in a tabulated data storage unit. [Figure 19] 11 is an example of a flowchart illustrating a procedure for a data counting process. [Figure 20] 13 is an example of a flowchart illustrating a procedure for a model determination process. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. First Embodiment Fig. 1 is a diagram showing an example of the configuration and processing of a lesion detection device according to a first embodiment. Lesion detection device 1 shown in Fig. 1 is an information processing device that detects a specific lesion area from a tomographic image of the inside of a human body. In the following description, as an example, it is assumed that the tomographic images are images taken at regular intervals in the height direction of the human body. This lesion detection device 1 has, for example, a storage unit 2 and a processing unit 3.

[0014] The storage unit 2 is a storage area secured in a storage device (not shown) included in the lesion detection device 1. The storage unit 2 stores data indicating the machine learning model 2a and data indicating the machine learning model 2b.

[0015] The machine learning model 2a is a trained model that identifies whether or not a pixel included in an input tomographic image is a specific lesion area. In the identification process using the machine learning model 2a, when a single tomographic image is input to the machine learning model 2a, a lesion area is identified from the tomographic image on a pixel-by-pixel basis.

[0016] On the other hand, the machine learning model 2b is a trained model that distinguishes whether or not an input tomographic image is a region of a specific lesion, using image blocks obtained by dividing the input tomographic image into fixed-size units. In the discrimination process using the machine learning model 2b, when a single tomographic image is given, the image is divided into image blocks. Then, the image blocks are input one by one to the machine learning model 2b, and each time, it is determined whether or not the input image block is a lesion region.

[0017] The processing unit 3 is, for example, a processor. In this case, the following processing of the processing unit 3 is realized by the processor executing a predetermined program. The processing unit 3 selectively uses either the machine learning model 2a or 2b to detect a lesion area from the captured tomographic image.

[0018] Here, the tomographic image group 12 is a collection of tomographic images including a diseased region among the tomographic images used as learning data when generating the machine learning model 2a by machine learning. The tomographic images included in the tomographic image group 12 are images obtained by photographing the inside of the human bodies 11a, 11b, ...

[0019] The processing unit 3 acquires a tomographic image 22a captured at a first imaging position inside the human body 21 as a target for lesion detection. Then, the processing unit 3 selects one of the machine learning models 2a and 2b based on the number of tomographic images captured at a second imaging position corresponding to the first imaging position in the human bodies 11a, 11b, etc., among the tomographic images included in the tomographic image group 12. The processing unit 3 detects a specific lesion area from the tomographic image 22a using the selected machine learning model.

[0020] The frequency of appearance of a lesion region may be biased depending on the position inside the human body. Therefore, the number of tomographic images including a lesion region among the tomographic images used as learning data may be biased depending on the position inside the human body. For tomographic images taken at positions where a sufficient number of tomographic images are not used as learning data, the accuracy of identifying lesions using the machine learning model 2a may be reduced.

[0021] In the above model selection, the processing unit 3 selects the machine learning model 2a when it is determined that the number of images of the tomographic images taken at the second shooting position is sufficient to obtain a certain level of lesion identification accuracy from the machine learning model 2a. On the other hand, the processing unit 3 selects the machine learning model 2b when it is determined that the number of images is not sufficient to obtain a certain level of lesion identification accuracy from the machine learning model 2a. Since the machine learning model 2b identifies the lesion area based on the features within the divided area rather than the entire tomographic image, the lesion identification accuracy does not depend on the shooting position of the tomographic image. Therefore, the above model selection makes it possible to detect the lesion area with a certain level of accuracy regardless of the shooting position of the tomographic image inside the human body.

[0022] In the example of FIG. 1, the lesion detection device 1 or another information processing device counts the number of tomographic images included in the tomographic image group 12 for each position in the human bodies 11a, 11b, .... Specifically, divided regions A1 to A3 are set by dividing a predetermined region of the human bodies 11a, 11b, ... into three at equal intervals in a direction perpendicular to the tomographic images. Then, the number of tomographic images is counted for each of the divided regions A1 to A3. In the example of FIG. 1, it is assumed that the tomographic image group 12 includes 1000 tomographic images, the number of images in the divided region A1 is 100, the number of images in the divided region A2 is 800, and the number of images in the divided region A3 is 100.

[0023] The divided regions A1 to A3 are obtained by dividing a region between predetermined reference points in each of the human bodies 11a, 11b, .... Specific organ regions can be applied as such regions.

[0024] In the example of FIG. 1, a group of tomographic images 22 is acquired by capturing images of a predetermined region of the human body 21. The predetermined region of the human body 21 is divided into three at equal intervals in a direction perpendicular to the tomographic image 22a to generate divided regions B1 to B3. Then, among the tomographic images included in the group of tomographic images 22, a lesion detection process is executed for the tomographic image 22a belonging to the divided region B2. In this case, the processing unit 3 acquires the number of images of the tomographic images used as learning data, which is tallied for the divided region A2 corresponding to the divided region B2, among the divided regions A1 to A3. Here, the number of images "800" is acquired for the divided region A2.

[0025] As an example, assume that 20% of the total number of tomographic images included in the tomographic image group 12 (1000 in FIG. 1) is set as the threshold. The ratio of the number of acquired images "800" to the total number "1000" is 80%, which is greater than the threshold, so the processing unit 3 selects the machine learning model 2a. In this case, the processing unit 3 inputs the tomographic image 22a to the machine learning model 2a and identifies whether or not each pixel of the tomographic image 22a is a lesion area.

[0026] On the other hand, for example, it is assumed that a lesion detection process is executed on a tomographic image belonging to division region B3 among the tomographic images included in the tomographic image group 22. In this case, the processing unit 3 acquires the number of images of the tomographic images used as learning data, which is tallied for division region A3 corresponding to division region B3 among division regions A1 to A3. Here, the number of images "100" is acquired for division region A3. Since the ratio of the acquired number of images "100" to the total number "1000" is 10%, which is less than the threshold value, the processing unit 3 selects the machine learning model 2b. In this case, the processing unit 3 divides the tomographic image into image blocks and inputs the divided image blocks one by one to the machine learning model 2b, thereby identifying whether or not each image block is a lesion region.

[0027] By such processing, it becomes possible to detect a lesion area with a certain degree of accuracy regardless of the position at which the tomographic image is taken inside the human body. Second Embodiment Next, we will explain a system that can detect lung lesion areas from CT images.

[0028] Fig. 2 is a diagram showing an example of the configuration of a diagnosis support system according to a second embodiment. The diagnosis support system shown in Fig. 2 is a system that supports image diagnosis work using CT imaging, and includes a CT device 50 and a diagnosis support device 100. Note that the diagnosis support device 100 is an example of the lesion detection device 1 shown in Fig. 1.

[0029] The CT device 50 captures X-ray CT images of the human body. In this embodiment, the CT device 50 captures a predetermined number of axial tomographic images of the chest region while changing the position (slice position) in the height direction of the human body (direction perpendicular to the axial plane) at predetermined intervals.

[0030] The diagnosis support device 100 extracts a lung field area from each tomographic image captured by the CT device 50, and detects a lesion area based on image information of the extracted lung field area. In this embodiment, a shadow area is detected as the lesion area. The diagnosis support device 100 may also identify the type of shadow. For example, the diagnosis support device 100 displays information indicating the detection result of the lesion area on a display device. In this way, the diagnosis support device 100 supports the image diagnosis work of a user (for example, a radiology doctor).

[0031] The diagnosis support device 100 detects a lesion area (or identifies a lesion) using a trained model that has been generated in advance by machine learning. The diagnosis support device 100 is also capable of executing a process of generating the trained model by machine learning.

[0032] Hereinafter, the hardware configuration of the diagnosis support device 100 will be described with reference to Fig. 2. The diagnosis support device 100 is realized, for example, as a computer as shown in Fig. 2. As shown in Fig. 2, the diagnosis support device 100 includes a processor 101, a random access memory (RAM) 102, a hard disk drive (HDD) 103, a graphics processing unit (GPU) 104, an input interface (I / F) 105, a reader 106, and a communication interface (I / F) 107.

[0033] The processor 101 comprehensively controls the entire diagnosis support device 100. The processor 101 is, for example, a central processing unit (CPU), a micro processing unit (MPU), a digital signal processor (DSP), an application specific integrated circuit (ASIC), or a programmable logic device (PLD). The processor 101 may also be a combination of two or more elements of a CPU, an MPU, a DSP, an ASIC, or a PLD. The processor 101 is an example of the processing unit 3 shown in FIG. 1.

[0034] The RAM 102 is used as a main storage device of the diagnosis support device 100. The RAM 102 temporarily stores at least a part of an OS (Operating System) program and application programs to be executed by the processor 101. The RAM 102 also stores various data necessary for processing by the processor 101.

[0035] The HDD 103 is used as an auxiliary storage device of the diagnosis support device 100. The OS program, application programs, and various data are stored in the HDD 103. Note that other types of non-volatile storage devices such as SSDs (Solid State Drives) can also be used as the auxiliary storage device.

[0036] A display device 104a is connected to the GPU 104. The GPU 104 displays an image on the display device 104a in accordance with an instruction from the processor 101. The display device 104a may be a liquid crystal display or an organic EL (Electroluminescence) display.

[0037] An input device 105a is connected to the input interface 105. The input interface 105 transmits a signal output from the input device 105a to the processor 101. The input device 105a may be a keyboard or a pointing device. The pointing device may be a mouse, a touch panel, a tablet, a touch pad, a trackball, or the like.

[0038] A portable recording medium 106a is detachably attached to the reading device 106. The reading device 106 reads data recorded on the portable recording medium 106a and transmits it to the processor 101. The portable recording medium 106a may be an optical disk, a semiconductor memory, or the like.

[0039] The communication interface 107 transmits and receives data to and from other devices such as the CT device 50 via the network. The processing functions of the diagnosis support device 100 can be realized by the above hardware configuration.

[0040] As described above, the process of identifying lesions from medical images can be performed using, for example, a trained model generated by machine learning. FIG. 3 is a diagram showing a first example of a trained model for lesion identification. Lesion identification model 60 shown in FIG. 3 is a trained model for identifying lesions by a segmentation method. Hereinafter, this lesion identification model 60 may be referred to as "segmentation model 60." Upon receiving an input of a tomographic image 62, segmentation model 60 identifies whether or not a particular lesion region is present (in this embodiment, whether or not a shadow is present) for each pixel of tomographic image 62. Here, it is assumed that tomographic image 62 is an axial plane tomographic image.

[0041] This segmentation model 60 is generated by machine learning (e.g., deep learning) using the axial plane tomographic image group 61 as training data. A label indicating whether or not each pixel is a lesion area is added to each tomographic image used as training data, and these labels are used as correct answer data during machine learning. Machine learning is performed on a tomographic image basis using such training data to generate the segmentation model 60.

[0042] The segmentation model 60 may be a lesion identification model that identifies three or more types of lesion types, including a type indicating no lesion, for each pixel. For example, the segmentation model 60 identifies, for each pixel, one of six types of shadow types, namely ground glass opacity, infiltrative opacity, honeycomb lung, emphysema, other shadows, and normal (no shadow). In this case, a label indicating which of the above shadow types each pixel is is added to each tomographic image used as learning data. Lesion identification processing using such a segmentation model 60 is effective for diagnosing, for example, diffuse pneumonia.

[0043] FIG. 4 is a diagram showing a second example of a trained model for lesion identification. Lesion identification model 70 shown in FIG. 4 is a trained model for identifying lesions by a classification method. Hereinafter, this lesion identification model 70 may be referred to as "Classification model 70". When a tomographic image 72 is input, Classification model 70 divides tomographic image 72 into patch images (image blocks) of a certain size, and identifies whether each divided patch image is a specific lesion area (in this embodiment, whether it is a shadow or not) as a unit. Note that, similar to tomographic image 62 in FIG. 3, tomographic image 72 is assumed to be a tomographic image of an axial plane.

[0044] This Classification model 70 is generated by machine learning (e.g., deep learning) using a group of axial plane tomographic images 71 as learning data. However, unlike the segmentation method of Fig. 3, each tomographic image used as learning data is divided into the above-mentioned patch images, and a label indicating whether or not it is a lesion area is added to each divided patch image. Then, each patch image is used as teacher data, and machine learning is performed with the patch image as a unit using the label as correct answer data, thereby generating the Classification model 70.

[0045] The classification model 70 may be a lesion classification model that classifies three or more lesion types, including a type indicating no lesion, for each patch image. For example, the classification model 70 classifies which of the above six shadow types each patch image belongs to. In this case, a label indicating which of the above shadow types each patch image belongs to is added to each tomographic image used as learning data.

[0046] In this way, the Classification model 70 identifies lesions by using patch images each containing a plurality of pixels as a unit. On the other hand, the Segmentation model 60 identifies lesions by pixel, and therefore can identify lesions by using smaller regions in a tomographic image as a unit, compared to the Classification model 70. However, as shown in the following Fig. 5, the Segmentation model 60 has a problem in that the identification accuracy may differ for each position of the tomographic image in the organ region.

[0047] FIG. 5 is a diagram showing the bias in the number of training data in organ regions. In generating segmentation model 60 by machine learning, among tomographic images obtained by photographing the chest region of a human body, tomographic images of organ regions (lung field regions in this embodiment) are used as training data. A training data set 81 shown in FIG. 5 is a collection of training data including, among such training data, tomographic images including a lesion region (a tomographic image including one or more pixels of a lesion region) as teacher data. In other words, the training data set 81 does not include a tomographic image without a lesion region. In the following description, the training data included in training data set 81 may be referred to as "training data of a lesion region".

[0048] The histogram 82 shown in Fig. 5 is a graph showing the number of pieces of learning data included in the learning dataset 81 for each divided region obtained by equally dividing the organ region in the vertical direction. As an example, the histogram 82 in Fig. 5 shows the number of pieces of learning data for each of divided regions R1 to R6 obtained by dividing the organ region into six regions from the top.

[0049] Here, where within an organ region a lesion region is likely to appear varies depending on the lesion to be identified. If there is a bias in the frequency of appearance of lesion regions depending on the position within an organ region, the number of learning data of lesion regions included in the above-mentioned learning data set 81 will also be biased depending on the position of the organ region. In the example of FIG. 5, lesion regions are likely to appear near the center in the vertical direction within the organ region. For this reason, in histogram 82, the number of learning data near the center in the vertical direction is greater than the upper and lower parts.

[0050] The segmentation model 60 identifies lesions pixel by pixel based on image features in the entirety of one tomographic image, such as the shape of an organ in the tomographic image and its positional relationship with other organs. In order to identify lesions with high accuracy, a sufficient number of learning data is required as learning data for the lesion area for each position in the organ area. However, if there is a bias in the number of learning data for each position in the organ area as in histogram 82, there is a problem that the accuracy of identifying lesions differs for each position in the organ area of ​​the tomographic image input to the segmentation model 60 during inference. In the example of histogram 82, when the upper and lower tomographic images in the vertical direction in the organ area are input, the accuracy of identifying lesions may decrease.

[0051] On the other hand, the Classification model 70 identifies lesions using patch images obtained by dividing a tomographic image as units. Since the Classification model 70 identifies lesions based on image features such as brightness of an image in a partial region of a tomographic image called a patch image, the classification accuracy does not depend on the position of the patch image in the organ region.

[0052] Therefore, diagnosis support device 100 of the present embodiment selects and uses one of segmentation model 60 and classification model 70 during inference depending on the number of data in the organ region of the learning data of the lesion region used when learning segmentation model 60. That is, when a tomographic image of a certain divided region in an organ region becomes an identification target, if the number of learning data of the lesion region in that divided region is large, lesion identification is performed using segmentation model 60. On the other hand, if the number of learning data of the lesion region in that divided region is small, lesion identification is performed using classification model 70. This makes it possible to suppress a decrease in the identification accuracy of lesions in divided regions with a small number of learning data of the lesion region, and to identify lesions with stable accuracy regardless of their position in the organ region.

[0053] Fig. 6 is a diagram showing an example of selection of a lesion identification model. As shown in Fig. 6, a threshold value TH is set for the number of pieces of learning data of a lesion region for each divided region. For example, the threshold value TH is set to a number of pieces of learning data that is a predetermined ratio (x%) of the total number of pieces of learning data included in learning dataset 81. In this case, when performing lesion identification for a tomographic image included in a certain divided region, if the number of pieces of learning data of the lesion region corresponding to the divided region is greater than threshold value TH, diagnosis support device 100 performs lesion identification using segmentation model 60. On the other hand, if the number of pieces of learning data of the lesion region is equal to or smaller than threshold value TH, diagnosis support device 100 performs lesion identification using classification model 70.

[0054] Fig. 7 is a diagram showing a specific example of model selection according to the number of learning data of a lesion region. In the example of Fig. 7, the organ region is divided vertically into three parts: "upper organ", "middle organ", and "lower organ". The total number of learning data included in the learning data set 81 is 1000, and the threshold value TH is set to 200, which corresponds to 20% of the total number.

[0055] 7, it is assumed that the number of pieces of learning data corresponding to the upper part of an organ among the pieces of learning data included in the learning data set 81 is 100. In this case, since the number of pieces of learning data is equal to or less than the threshold value TH, lesion classification for the tomographic images corresponding to the upper part of an organ is performed using the Classification model 70.

[0056] Also, it is assumed that the number of pieces of training data corresponding to the central part of the organ among the training data set 81 is 800. In this case, since the number of pieces of training data is greater than the threshold value TH, lesion classification for the tomographic images corresponding to the central part of the organ is performed using the segmentation model 60.

[0057] Furthermore, it is assumed that the number of pieces of learning data corresponding to the lower part of the organ among the pieces of learning data included in the learning data set 81 is 100. In this case, since the number of pieces of learning data is equal to or less than the threshold value TH, lesion classification for the tomographic images corresponding to the upper part of the organ is performed using the Classification model 70.

[0058] This process can suppress a decrease in the accuracy of lesion identification in divided regions where the number of learning data for the lesion region is small. As a result, lesions can be identified with stable accuracy regardless of their position in the organ region of the tomographic image to be identified.

[0059] The threshold value TH may be set as, for example, the absolute number of pieces of learning data itself, in addition to the value indicating the ratio to the total number of pieces of learning data as in the examples of Figures 6 and 7. For example, depending on the lesion to be identified and the corresponding disease, if a certain number or more pieces of learning data of the lesion area are secured, a certain level of identification accuracy or higher may be obtained. In such a case, the threshold value TH may be the minimum number of pieces of learning data of the lesion area required to obtain a certain level of identification accuracy or higher.

[0060] In addition, the division direction of the organ region for generating the divided regions is not limited to the vertical direction as described above. For example, the organ region may be divided evenly in the direction perpendicular to the slice plane of the tomographic image (parallel direction of the tomographic image).

[0061] 8 is a diagram showing an example of the configuration of processing functions of the diagnosis support device 100. The diagnosis support device 100 includes a learning data storage unit 110, a setting data storage unit 120, a compilation data storage unit 130, a model data storage unit 140, a data compilation unit 151, a model determination unit 152, a model generation unit 153, an organ region extraction unit 154, a model selection unit 155, and a lesion identification unit 156.

[0062] The learning data storage unit 110, the setting data storage unit 120, the collected data storage unit 130, and the model data storage unit 140 are storage areas secured in a storage device included in the diagnosis support device 100, such as the RAM 102 or the HDD 103.

[0063] The learning data storage unit 110 stores learning data used when generating the segmentation model 60 and the classification model 70 by machine learning. The setting data storage unit 120 stores setting data that is referenced in the process of determining the lesion identification model to be used for each divided region.

[0064] The aggregated data storage unit 130 stores aggregated data on the learning data of the lesion region (learning data including tomographic images including the lesion region as teacher data) from among the learning data stored in the learning data storage unit 110. The aggregated data stores the number of pieces of learning data of the lesion region for each divided region obtained by dividing the organ region. The aggregated data storage unit 130 also stores information indicating whether the segmentation model 60 or the classification model 70 is to be used for each divided region.

[0065] The model data storage unit 140 stores model parameters corresponding to an organ region extraction model, which is a trained model for extracting organ regions, the segmentation model 60, and the classification model 70.

[0066] The processes of the data aggregation unit 151, the model determination unit 152, the model generation unit 153, the organ region extraction unit 154, the model selection unit 155 and the lesion identification unit 156 are realized, for example, by the processor 101 executing a predetermined program.

[0067] The data aggregation unit 151, the model determination unit 152, and the model generation unit 153 operate in a learning phase in which the segmentation model 60 and the classification model 70 are generated by machine learning.

[0068] Data counting unit 151 counts the number of pieces of learning data of the lesion area for each divided area out of the learning data stored in learning data storage unit 110, and registers the counting result in collected data storage unit .

[0069] The model determination unit 152 determines whether to use the segmentation model 60 or the classification model 70 for each divided region based on the number of learning data for each divided region registered in the aggregated data storage unit 130, and registers the result in the aggregated data storage unit 130.

[0070] The model generation unit 153 generates a segmentation model 60 and a classification model 70 by machine learning using the learning data stored in the learning data storage unit 110, and registers model parameters indicating each of the generated lesion identification models in the model data storage unit 140.

[0071] The organ region extraction unit 154, the model selection unit 155 and the lesion identification unit 156 operate in an inference phase to identify lesions using the segmentation model 60 and the classification model 70.

[0072] The organ region extraction unit 154 extracts, from a set of tomographic images obtained by imaging the diagnostic subject, a tomographic image included in the organ region, using an organ region extraction model. The model selection unit 155 selects either the segmentation model 60 or the classification model 70 as the lesion identification model to be used when identifying a lesion, for each of the cross-sectional images extracted by the organ area extraction unit 154, based on the information registered in the aggregated data storage unit 130.

[0073] Lesion identifying section 156 uses the lesion identification model selected by model selecting section 155 to identify whether or not a region is a lesion area from the tomographic image on a pixel-by-pixel or patch-image basis. 9 is a diagram showing an example of data stored in the learning data storage unit 110. In the learning data storage unit 110, learning data tables 111 and 112 are stored.

[0074] The learning data table 111 stores learning data used to generate the segmentation model 60. The learning data table 111 stores multiple sets of learning data including image data and correct answer data for each of multiple cases. The image data is data of tomographic images, and each image data group corresponding to a case is an image data group of multiple tomographic images obtained by capturing an image of one person once. The correct answer data is data in which a label indicating whether or not each pixel of the corresponding tomographic image is associated with a pixel.

[0075] The learning data table 112 registers learning data used for generating the Classification model 70. The learning data table 112 registers multiple sets of learning data including image data, which is teacher data, and correct answer data for each of multiple cases. The image data is data of patch images obtained by dividing a tomographic image into a predetermined size, and the image data group corresponding to a case is an image data group of patch images obtained from each of multiple tomographic images obtained by one shooting of one person. The correct answer data is a label indicating whether the corresponding patch image is a lesion or not.

[0076] 10 is a diagram showing an example of data stored in the setting data storage unit. As described above, the setting data storage unit 120 stores setting data that is referenced in the process of determining the lesion identification model to be used for each divided region. The setting data storage unit 120 stores, as such data, a setting data set 121 including the number of divisions and a threshold value.

[0077] The number of divisions indicates the number of divisions into which the organ region is divided (i.e., the number of divided regions into which the organ region is divided). The threshold value indicates the threshold value (the aforementioned threshold value TH) used when deciding which lesion identification model to use, expressed as a ratio to the total number of learning data for the lesion region. The number of divisions and the threshold value can be set arbitrarily by the user.

[0078] 11 is a diagram showing an example of data stored in the aggregate data storage unit 130. The aggregate data storage unit 130 stores an aggregate data table 131. A record for each divided area is registered in the aggregate data storage unit 130. Each record includes a divided area range, the number of pieces of learning data, and a model flag.

[0079] The divided area range is a value that indicates the range of the divided area in the organ area as a ratio based on the upper end of the organ area. In the example of Fig. 11, the range of the divided area is indicated by a ratio in which the upper end of the organ area is 0 and the lower end is 1. When the organ area is divided into three areas as in Fig. 11, for example, the divided area range of the first divided area is expressed as 0 or more and 0.33 or less, the divided area range of the second divided area is expressed as greater than 0.33 and less than 0.66, and the divided area range of the third divided area is expressed as greater than 0.66 and less than 1.0.

[0080] The number of pieces of training data indicates the number of pieces of training data including tomographic images included in the corresponding divided regions, among the training data of the lesion region, used when generating the segmentation model 60 . The model flag is a flag indicating whether to use the Segmentation model 60 or the Classification model 70 when identifying a lesion in a tomographic image included in the corresponding divided region. In this embodiment, the flag is set to 0 when the Segmentation model 60 is used, and the flag is set to 1 when the Classification model 70 is used.

[0081] 12 is a diagram showing an example of data stored in the model data storage unit 140. A model data table 141 is stored in the model data storage unit 140. In the model data table 141, model structure data and weight data are registered for each of the organ region extraction model, the segmentation model 60, and the classification model 70.

[0082] The model structure data indicates the type of machine learning model used and the configuration of the neural network. In reality, the file name of a file describing such information may be registered in the model data table 141. The weight data indicates a weight coefficient set between nodes in the neural network of the machine learning model. In reality, the model data table 141 may indicate the file name of a file describing such weight coefficients.

[0083] The model structure data and weight data for the organ region extraction model are registered in advance in the model data table 141. The model structure data for the segmentation model 60 and the classification model 70 are also registered in advance in the model data table 141. Meanwhile, the weight data for the segmentation model 60 and the classification model 70 are generated by the model generation unit 153 and registered in the model data table 141.

[0084] Next, a description will be given with reference to a flowchart of the processing performed by the diagnosis support device 100. First, the processing performed in the learning phase will be described with reference to Figs. FIG. 13 is an example of a flowchart illustrating a procedure of the data counting process.

[0085] [Step S11] The data aggregation unit 151 acquires the number of divisions of the organ region from the setting data set 121 in the setting data storage unit 120. [Step S12] The data aggregation unit 151 creates records in the aggregated data table 131 of the aggregated data storage unit 130 for the number of divisions obtained, and sets the range of each divided region (divided region range). If the number of divisions is d, the divided region ranges are, in order from the top of the body, from 0 to 1 / d, from 1 / d to 2 / d, ..., from d-1 / d to 1. The data aggregation unit 151 also sets the number of learning data for each record created in the aggregated data table 131 to an initial value of 0.

[0086] [Step S13 ] The data aggregation unit 151 selects one case from the learning data table 111 for the segmentation model 60 stored in the learning data storage unit 110 . Here, the data aggregation unit 151 assigns image numbers, starting from 0, to the image data of the tomographic images included in the selected case, in order from the image data of the upper end of the body. Then, the data aggregation unit 151 assigns the image number of the image data of the upper end of the organ region among these image data, as I min Let I be the image number of the image data at the bottom of the organ region. max Let us assume that.

[0087] Since all the tomographic images registered as learning data are images within the organ region, in step S13, min =0, I max =I min It is set to +(the number of image data included in the selected case)-1.

[0088] [Step S14] The data aggregation unit 151 selects one image data included in the selected case. In this process, one unselected image data is selected from the image data included in the case, starting from the top (top).

[0089] [Step S15] Data aggregation unit 151 determines whether the selected image data includes a lesion area. In this process, data aggregation unit 151 acquires supervised data corresponding to the selected image data from learning data table 111. Data aggregation unit 151 determines that the selected image data includes a lesion area when the acquired supervised data has one or more pixels labeled with a lesion. If it is determined that the selected image data includes a lesion area, processing proceeds to step S16; if it is determined that the selected image data does not include a lesion area, processing proceeds to step S18.

[0090] [Step S16] In order to determine which divided area the selected image data is included in, the data aggregation unit 151 calculates the normalized image number x of the selected image data. afterIf the image number of the selected image data is x, then image number x after is calculated by the following formula (1). x after =(xI min ) / (I max -I min ) ···(1) According to the formula (1), the image number x is normalized to a value between 0 and 1. That is, the image number x calculated according to the formula (1) after indicates the percentage of the position of the selected image data from the leading edge of the entire organ region.

[0091] [Step S17] The data aggregation unit 151 calculates the calculated image number x after In step S12, the data aggregation unit 151 determines which divided area range set in the aggregated data table 131 the selected image data belongs to. In this way, the data aggregation unit 151 determines the divided area to which the selected image data belongs. The data aggregation unit 151 counts up by 1 the number of pieces of learning data in the aggregated data table 131 that corresponds to the divided area to which the selected image data belongs.

[0092] [Step S18] The data aggregation unit 151 determines whether all image data included in the case selected in step S13 have been selected. If there is unselected image data, the process proceeds to step S14, where the image data at the top of the unselected image data is selected. On the other hand, if all the corresponding image data have been selected, the process proceeds to step S19.

[0093] [Step S19] The data aggregation unit 151 determines whether all cases included in the learning data table 111 have been selected. If there are unselected cases, the process proceeds to step S13, where one of the unselected cases is selected. On the other hand, if all cases have been selected, the data aggregation process ends.

[0094] Through the above process, the number of pieces of learning data for each divided region is registered in the aggregated data table 131. Then, with reference to such aggregated data table 131, the model determination process shown in FIG.

[0095] FIG. 14 is an example of a flowchart illustrating a procedure of the model determination process. [Step S21] The model determination unit 152 acquires a threshold value from the setting data set 121 in the setting data storage unit 120.

[0096] [Step S22] The model determination unit 152 selects one of the divided regions whose divided region ranges are set in the aggregated data table 131 in the aggregated data storage unit 130. [Step S23] The model determination unit 152 obtains from the aggregated data table 131 the number of pieces of learning data corresponding to the selected divided region.

[0097] [Step S24] The model determination unit 152 determines whether the number of pieces of training data acquired is equal to or less than the threshold acquired in step S21. If the number of pieces of training data is equal to or less than the threshold, the process proceeds to step S25. If the number of pieces of training data is greater than the threshold, the process proceeds to step S26.

[0098] If the threshold registered in the setting data set 121 is a value indicating a ratio, the model determination unit 152 converts this threshold into an absolute number of learning data and compares the converted threshold with the acquired number of learning data. Specifically, the model determination unit 152 sums up the number of learning data for each divided area registered in the aggregated data table 131 and multiplies the total value by a threshold indicating a ratio to convert it into a threshold indicating the absolute number of learning data.

[0099] [Step S25] The model determination unit 152 sets 1 as a model flag corresponding to the selected divided region in the aggregated data table 131. This sets that the Classification model 70 is to be used for lesion classification of the tomographic image included in the selected divided region.

[0100] [Step S26] The model determination unit 152 sets 0 as a model flag corresponding to the selected divided region in the aggregated data table 131. This sets that the segmentation model 60 is to be used for lesion identification of the tomographic image included in the selected divided region.

[0101] [Step S27] The model determination unit 152 determines whether all divided regions have been selected. If there are any unselected divided regions, the process proceeds to step S22, where one of the unselected divided regions is selected. On the other hand, if all divided regions have been selected, the model determination process ends.

[0102] FIG. 15 is an example of a flowchart illustrating a procedure of a model generation process. [Step S31] The model generation unit 153 obtains training data for the segmentation model 60 from the training data table 111.

[0103] [Step S32] The model generation unit 153 executes machine learning using the acquired learning data to generate the segmentation model 60. Specifically, the model generation unit 153 calculates weight data for the segmentation model 60 by machine learning, and registers the weight data in the record corresponding to the segmentation model 60 among the records included in the model data table 141.

[0104] [Step S33] The model generation unit 153 obtains training data for the Classification model 70 from the training data table 112. [Step S34] The model generation unit 153 executes machine learning using the acquired learning data to generate the Classification model 70. Specifically, the model generation unit 153 calculates weight data for the Classification model 70 by machine learning, and registers the weight data in the record corresponding to the Classification model 70 among the records included in the model data table 141.

[0105] The model generation process in FIG. 15 may be executed after the processes in FIGS. 13 and 14, or may be executed before the processes in FIGS. Next, the processing in the inference phase will be described with reference to FIG.

[0106] FIG. 16 is an example of a flowchart showing the procedure of the lesion identification process. [Step S41] The organ region extraction unit 154 acquires a set of cross-sectional images of the subject captured by the CT apparatus 50. The organ region extraction unit 154 assigns image numbers, starting from 1, to each cross-sectional image included in the acquired cross-sectional image set, in order from the top of the body.

[0107] [Step S42] The organ region extraction unit 154 inputs the acquired set of tomographic images to an organ region extraction model based on weighting data registered in the model data table 141. As a result, the organ region extraction unit 154 extracts tomographic images included in the organ region from the tomographic images included in the set of tomographic images.

[0108] The model selection unit 155 assigns the image number of the first (upper end) tomographic image among the extracted tomographic images as I min The image number of the last (bottom) tomographic image is I max Let us assume that. [Step S43] The model selection unit 155 selects one cross-sectional image of the organ region. In this process, unselected cross-sectional images are selected from the cross-sectional images extracted in step S42, starting from the top.

[0109] [Step S44] In order to determine which divided region contains the selected tomographic image, the model selection unit 155 calculates the normalized image number x after If the image number of the selected tomographic image is x, then image number x after is calculated by the above-mentioned formula (1). This allows calculation of the percentage of the position of the selected tomographic image from the leading edge of the entire organ region.

[0110] [Step S45] The model selection unit 155 selects the calculated image number x from among the divided region ranges registered in the tabulated data table 131. after In this way, the model selection unit 155 determines the divided area to which the selected tomographic image belongs.

[0111] The model selection unit 155 acquires a model flag corresponding to the divided area to which the selected tomographic image belongs from among the model flags in the aggregated data table 131, and determines whether the acquired model flag is 1. If the model flag is 0, the process proceeds to step S46, and if the model flag is 1, the process proceeds to step S47.

[0112] [Step S46] Lesion identifier 156 inputs the tomographic image selected in step S43 to segmentation model 60, which is based on the weighting data registered in model data table 141. In this way, lesion identifier 156 identifies whether or not each pixel on the tomographic image represents a lesion.

[0113] [Step S47] The lesion identifying unit 156 divides the tomographic image selected in step S43 into patch images of a certain size. [Step S48] Lesion classifier 156 inputs the divided patch images one by one to classification model 70, which is based on the weighting data registered in model data table 141. In this way, lesion classifier 156 classifies each patch image on the tomographic image as whether or not it is a lesion.

[0114] [Step S49] The lesion identification unit 156 determines whether all of the cross-sectional images in the organ region have been selected. If there are unselected cross-sectional images, the process proceeds to step S43, where one of the unselected cross-sectional images is selected. On the other hand, if all of the corresponding cross-sectional images have been selected, the lesion identification process ends.

[0115] In steps S46 and S48, lesion identification unit 156 generates a classification result image in which, for example, information indicating whether or not each pixel of the selected tomographic image is a lesion is added. When a request is subsequently made to display the classification result of the tomographic image, lesion identification unit 156 displays the area determined to be a lesion on the tomographic image, distinguishing it from other areas, based on the classification result image.

[0116] As described above, the segmentation model 60 and the classification model 70 may be lesion classification models that classify three or more types of lesion types including a type indicating no lesion. In this case, in steps S46 and S48, for example, a classification result image is generated in which information indicating one of the lesion types is assigned to each pixel of the tomographic image.

[0117] The above process can suppress a decrease in the accuracy of lesion identification in divided regions where the number of learning data for the lesion region is small. As a result, lesions can be identified with stable accuracy regardless of their position in the organ region of the tomographic image to be identified.

[0118] 13 to 15 in the learning phase and the process of FIG. 16 in the inference phase may be executed in different information processing devices. In this case, for example, the learning data storage unit 110, the setting data storage unit 120, the tabulated data storage unit 130, the model data storage unit 140, the data tabulating unit 151, the model determining unit 152, and the model generating unit 153 are implemented in a first information processing device. On the other hand, the organ region extracting unit 154, the model selecting unit 155, and the lesion identifying unit 156 are implemented in a second information processing device. Then, the data stored in the setting data storage unit 120, the tabulated data storage unit 130, and the model data storage unit 140 are transferred to the second information processing device via a network or a portable recording medium, and these data are referred to in the process of the inference phase.

[0119] Third embodiment In the above-described second embodiment, a lesion identification model is selected for a division region to which a captured tomographic image belongs, depending on whether or not the number of pieces of learning data for the lesion region used during learning is large. In contrast, in the third embodiment, when segmentation model 60 is generated, the identification accuracy of segmentation model 60 is verified for each division region. Then, in the inference phase, a lesion identification model is selected for a division region to which a captured tomographic image belongs, depending on whether or not the identification accuracy of segmentation model 60 is high.

[0120] Fig. 17 is a diagram showing an example of the configuration of processing functions included in a diagnosis support device according to a third embodiment. In Fig. 17, the same components as those in Fig. 8 are denoted by the same reference numerals, and the description thereof will be omitted.

[0121] A diagnosis support device 100a according to the third embodiment includes a learning data storage unit 110a, a setting data storage unit 120a, and a total data storage unit 130a, instead of the learning data storage unit 110, the setting data storage unit 120, and the total data storage unit 130 in FIG.

[0122] The learning data storage unit 110a, like the learning data storage unit 110, includes learning data used when generating the segmentation model 60 and learning data used when generating the classification model 70. However, a part of the former learning data is used to verify the classification accuracy of the segmentation model 60.

[0123] The setting data storage unit 120a differs from the setting data storage unit 120 in that a value of the discrimination accuracy is set as a threshold value. In the aggregate data storage unit 130a, the number of verifications and the number of correct answers when verifying the classification accuracy of the segmentation model 60 are registered instead of the number of pieces of training data.

[0124] 8. Moreover, the diagnosis support device 100a includes a data collection unit 151a and a model determination unit 152a instead of the data collection unit 151 and the model determination unit 152 of FIG. After the segmentation model 60 is generated by the model generation unit 153, the data aggregation unit 151a verifies the classification accuracy of the segmentation model 60 for each divided region by using the learning data for verification registered in the learning data storage unit 110a.

[0125] The model determining unit 152a determines a lesion identification model to be used for each divided region based on the verification result by the data aggregating unit 151a. 18 is a diagram showing an example of data stored in the aggregate data storage unit. The aggregate data storage unit 130a stores an aggregate data table 132. A record for each divided area is registered in the aggregate data table 132. Each record includes a divided area range, a verification count, a correct answer count, and a model flag.

[0126] The divided region range and the model flag are the same as those in the tabulated data table 131 shown in Fig. 11. The number of verifications indicates the number of tomographic images (or the number of pixels) whose classification accuracy has been verified using the segmentation model 60. The number of correct answers indicates the number of tomographic images (or the number of pixels) that have been determined to be correct by verification.

[0127] FIG. 19 is an example of a flowchart illustrating a procedure of the data counting process. The data aggregation unit 151a executes the processes of steps S11 and S12 in Fig. 13. In step S12, a record is created in the aggregated data table 132. Next, the processes of the next steps S51 and onwards are executed.

[0128] [Step S51] The data aggregation unit 151a selects one verification case from the learning data table for the segmentation model 60 stored in the learning data storage unit 110a. Here, the data aggregation unit 151 assigns image numbers, starting from 0, to the image data of the tomographic images included in the selected case, in order from the image data of the upper end of the body. Then, the data aggregation unit 151 assigns the image number of the image data of the upper end of the organ region among these image data, as I minLet I be the image number of the image data at the bottom of the organ region. max Let us assume that.

[0129] Since all the tomographic images registered as learning data are images within the organ region, in step S51, min =0, I max =I min It is set to +(the number of image data included in the selected case)-1.

[0130] [Step S52] The data aggregation unit 151a selects one image data included in the selected case. In this process, one unselected image data is selected from the image data included in the case, starting from the top (upper side).

[0131] [Step S53] In order to determine which divided area the selected image data is included in, the data aggregation unit 151a calculates the normalized image number x of the selected image data. after If the image number of the selected image data is x, then image number x after is calculated by the above-mentioned formula (1). This allows the calculation of the percentage of the position of the selected image data from the leading edge of the entire organ region.

[0132] The data aggregation unit 151a calculates the image number x after In step S12, the data aggregation unit 151a determines which divided area range the selected image data belongs to. The data aggregation unit 151a counts up by 1 the verification number in the aggregated data table 132 that corresponds to the divided area to which the selected image data belongs.

[0133] [Step S54] The data aggregation unit 151a inputs the image data selected in step S52 to the segmentation model 60 based on the weighting data registered in the model data table 141. As a result, the data aggregation unit 151a identifies whether or not each pixel on the tomographic image represents a lesion.

[0134] [Step S55] The data aggregation unit 151a compares the classification result for each pixel in step S54 with the value for each pixel of the correct answer data corresponding to the selected image data. If the classification result and the pixel value match, the classification is deemed to be correct. For example, if the number of correct answers among the pixels of the image data is equal to or greater than a certain percentage, the data aggregation unit 151a determines that the classification accuracy for the selected image data is high. In this case, the data aggregation unit 151a counts up by one the number of correct answers in the aggregate data table 132 that corresponds to the divided area to which the selected image data belongs.

[0135] In step S53, the number of pixels of the image data may be added to the corresponding verification number registered in the aggregated data table 132. In this case, in step S55, the number of pixels that were correctly answered may be added to the corresponding correct answer number registered in the aggregated data table 132.

[0136] [Step S56] The data aggregation unit 151a determines whether all image data included in the case selected in step S51 have been selected. If there is unselected image data, the process proceeds to step S52, where the image data at the top of the unselected image data is selected. On the other hand, if all the corresponding image data have been selected, the process proceeds to step S57.

[0137] [Step S57] The data aggregation unit 151a determines whether all the cases for verification included in the learning data table 111 have been selected. If there are unselected cases, the process proceeds to step S51, where one of the unselected cases is selected. On the other hand, if all the cases have been selected, the data aggregation process ends.

[0138] Through the above process, the verification number and the correct answer number for each divided region are registered in the tabulated data table 132. Then, with reference to such tabulated data table 132, the model determination process shown in FIG.

[0139] FIG. 20 is an example of a flowchart illustrating a procedure for a model determination process. In the process of Fig. 20, the processes of steps S21, S22, and S24 to S27 in Fig. 14 are executed by the model determination unit 152a. Moreover, instead of step S23 in Fig. 14, the following step S23a is executed.

[0140] [Step S23a] The model determination unit 152a obtains the number of verifications and the number of correct answers from a record corresponding to the selected divided area among the records in the aggregated data table 132. The model determination unit 152a calculates the classification accuracy corresponding to the selected divided area based on the number of correct answers / the number of verifications.

[0141] In the next step S24, the calculated discrimination accuracy is compared with a threshold value. According to the third embodiment described above, it becomes possible to identify a lesion with stable accuracy, regardless of its position in an organ region of a tomographic image to be identified.

[0142] The processing functions of the devices (e.g., lesion detection device 1, diagnosis support devices 100, 100a) shown in each of the above embodiments can be realized by a computer. In this case, a program describing the processing contents of the functions that each device should have is provided, and the above processing functions are realized on the computer by executing the program on a computer. The program describing the processing contents can be recorded on a computer-readable recording medium. Examples of computer-readable recording media include magnetic storage devices, optical discs, and semiconductor memories. Examples of magnetic storage devices include hard disk drives (HDDs) and magnetic tapes. Examples of optical discs include CDs (Compact Discs), DVDs (Digital Versatile Discs), and Blu-ray Discs (BD, registered trademark).

[0143] When distributing a program, for example, the program is recorded on a portable recording medium such as a DVD or CD and then sold. The program can also be stored in a storage device of a server computer and transferred from the server computer to other computers via a network.

[0144] A computer that executes a program stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. The computer then reads the program from its own storage device and executes processing according to the program. The computer can also read the program directly from a portable recording medium and execute processing according to the program. The computer can also execute processing according to the received program each time a program is transferred from a server computer connected via a network. [Explanation of symbols]

[0145] 1 Lesion detection device 2 Storage section 2a,2b Machine learning models 3 Processing section 11a,11b,21 Human body 12,22 Tomographic images 22a Tomographic image A1~A3,B1~B3 divided area

Claims

1. The computer acquiring a first tomographic image captured at a first imaging position inside a first human body; a first machine learning model that identifies whether or not a region of a specific lesion is included in a plurality of tomographic images taken inside one or more second human bodies, the first machine learning model being used as learning data when a first machine learning model that identifies whether or not a region of a specific lesion is included in a unit of pixels included in an input tomographic image is generated by machine learning, based on a number of tomographic images taken at a second imaging position corresponding to the first imaging position inside the one or more second human bodies, the first machine learning model or a second machine learning model that identifies whether or not a region of a specific lesion is included in a unit of image blocks obtained by dividing the input tomographic image into image blocks of a certain size; detecting a region of the specific lesion from the first tomographic image using a machine learning model selected from the first machine learning model and the second machine learning model; Lesion detection methods.

2. The computer further executes a process of counting the number of tomographic images captured in a corresponding first divided area among the plurality of tomographic images for each of a plurality of first divided areas obtained by dividing a predetermined organ area in the one or more second human bodies in a predetermined direction; In the selection, Identifying a third divided area in which the first tomographic image is captured, among a plurality of second divided areas obtained by dividing the organ area in the first human body into the predetermined number of divisions in the predetermined direction; selecting one of the first machine learning model and the second machine learning model based on the number of images counted for a fourth divided area corresponding to the third divided area among the plurality of first divided areas; The method of claim 1 .

3. In the selection, the first machine learning model is selected when a ratio of the number of images counted for the fourth divided region to a total number of the plurality of tomographic images is greater than a predetermined threshold, and the second machine learning model is selected when the ratio is equal to or less than the predetermined threshold. The method of claim 2 .

4. The computer acquiring a plurality of tomographic images obtained by photographing the inside of one or more second human bodies, and a plurality of pieces of correct answer data corresponding to each of the plurality of tomographic images and indicating whether or not each pixel on the corresponding tomographic image is a region of a specific lesion; classifying the plurality of tomographic images into any one of a plurality of first divided regions obtained by dividing a predetermined organ region in the one or more second human bodies into a predetermined number of divisions in a predetermined direction, depending on which divided region the plurality of tomographic images were taken in; a classification process is executed by inputting the plurality of tomographic images into a first machine learning model that classifies whether or not a pixel included in the input tomographic image is a specific lesion region, and a classification result is compared with the plurality of correct answer data to calculate an index indicating the classification accuracy of the first machine learning model for each of the plurality of first divided regions; acquiring a first tomographic image of the inside of a first human body; Identifying a third divided area in which the first tomographic image is captured, among a plurality of second divided areas obtained by dividing the organ area in the first human body into the predetermined number of divisions in the predetermined direction; selecting, based on the index calculated for a fourth divided region corresponding to the third divided region among the plurality of first divided regions, either the first machine learning model or a second machine learning model that identifies whether or not an image block obtained by dividing the input tomographic image into a certain size is a region of the specific lesion; detecting a region of the specific lesion from the first tomographic image using a machine learning model selected from the first machine learning model and the second machine learning model; Lesion detection methods.

5. In the selection, the first machine learning model is selected when the index calculated for the fourth divided region is greater than a predetermined threshold, and the second machine learning model is selected when the index is equal to or less than the predetermined threshold. The method for detecting a lesion according to claim 4.

6. On the computer, acquiring a first tomographic image captured at a first imaging position inside a first human body; a first machine learning model that identifies whether or not a region of a specific lesion is included in a plurality of tomographic images taken inside one or more second human bodies, the first machine learning model being used as learning data when a first machine learning model that identifies whether or not a region of a specific lesion is included in a unit of pixels included in an input tomographic image is generated by machine learning, based on a number of tomographic images taken at a second imaging position corresponding to the first imaging position inside the one or more second human bodies, the first machine learning model or a second machine learning model that identifies whether or not a region of a specific lesion is included in a unit of image blocks obtained by dividing the input tomographic image into image blocks of a certain size; detecting a region of the specific lesion from the first tomographic image using a machine learning model selected from the first machine learning model and the second machine learning model; A lesion detection program to carry out the process.

7. On the computer, acquiring a plurality of tomographic images obtained by photographing the inside of one or more second human bodies, and a plurality of pieces of correct answer data corresponding to each of the plurality of tomographic images and indicating whether or not each pixel on the corresponding tomographic image is a region of a specific lesion; classifying the plurality of tomographic images into any one of a plurality of first divided regions obtained by dividing a predetermined organ region in the one or more second human bodies into a predetermined number of divisions in a predetermined direction, depending on which divided region the plurality of tomographic images were taken in; a classification process is executed by inputting the plurality of tomographic images into a first machine learning model that classifies whether or not a pixel included in the input tomographic image is a specific lesion region, and a classification result is compared with the plurality of correct answer data to calculate an index indicating the classification accuracy of the first machine learning model for each of the plurality of first divided regions; acquiring a first tomographic image of the inside of a first human body; Identifying a third divided area in which the first tomographic image is captured, among a plurality of second divided areas obtained by dividing the organ area in the first human body into the predetermined number of divisions in the predetermined direction; selecting, based on the index calculated for a fourth divided region corresponding to the third divided region among the plurality of first divided regions, either the first machine learning model or a second machine learning model that identifies whether or not an image block obtained by dividing the input tomographic image into a certain size is a region of the specific lesion; detecting a region of the specific lesion from the first tomographic image using a machine learning model selected from the first machine learning model and the second machine learning model; A lesion detection program to carry out the process.

Images