Lesion detection method and lesion detection program

Abstract

To improve the detection accuracy of a lesion area from a medical image.SOLUTION: A computer classifies multiple first tomographic images into multiple first tomographic image groups having different medical findings, and generates multiple lesion identification models for identifying whether or not each of unit image regions is a region of a specific lesion by machine learning using one of mutually different ones of the multiple first tomographic image groups as learning data. The computer also calculates a first image feature amount on the basis of multiple second tomographic images, inputs the multiple second tomographic images to each of the multiple lesion identification models to obtain a probability that each of the unit image regions included in the multiple second tomographic images is the region of the specific lesion from each of the multiple lesion identification models, and integrates the probabilities obtained from each of the multiple lesion identification models on the basis of the first image feature amount for each of the unit image region included in the multiple second tomographic images.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 technique is the detection of lesion regions in medical images using a trained model created by machine learning. For example, an artificial intelligence pipeline for lesion detection and classification has been proposed that includes multiple trained machine learning models. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] US Patent Application Publication No. 2022 / 0270254 Summary of the Invention [Problem to be solved by the invention]

[0005] Incidentally, in a lesion area detection process using a trained model, the greater the difference in pixel values ​​(e.g., luminance values) between a lesion area and a non-lesion normal area, the higher the detection accuracy can be. However, in actual medical images, the difference in pixel values ​​between a lesion area and a normal area is often small, which can make it difficult to improve the detection accuracy. In addition, when a medical event occurs that can affect pixel values, such as fat accumulated in an organ, the difference in pixel values ​​between the lesion area and the normal area may be smaller than usual. Furthermore, in some cases, the magnitude relationship between the pixel values ​​may be reversed.

[0006] According to one aspect, the present invention aims to provide a lesion detection method and a lesion detection program capable of improving the accuracy of detecting a lesion area from a medical image. [Means for solving the problem]

[0007] In one proposal, a lesion detection method is provided in which a computer executes a learning process and a lesion detection process. In the learning process, the computer classifies a plurality of first tomographic images obtained by photographing the inside of a plurality of first human bodies into a plurality of first tomographic image groups having different medical findings, and generates a plurality of first lesion identification models that identify whether each unit image area included in the tomographic image to be identified is a region of a specific lesion by machine learning using one of the plurality of first tomographic image groups that is different from each other as training data. In the lesion detection process, the computer calculates a first image feature amount based on a plurality of second tomographic images obtained by photographing the inside of a second human body, inputs the plurality of second tomographic images to each of the plurality of first lesion identification models, and obtains a probability that each unit image area included in the plurality of second tomographic images is a region of a specific lesion from each of the plurality of first lesion identification models, integrates the probabilities obtained from each of the plurality of first lesion identification models for each unit image area included in the plurality of second tomographic images based on the first image feature amount to calculate an integrated value, and detects a region of a specific lesion from each of the plurality of second tomographic images based on the integrated value.

[0008] In addition, in one proposal, 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

[0009] In one aspect, the accuracy of detecting a lesion area from a medical image is improved. [Brief description of the drawings]

[0010] [Figure 1] 1 illustrates a configuration example and a processing example of an information processing 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 an example of a machine learning model for detection of lesion areas. [Figure 4] FIG. 2 is a diagram illustrating an example of the configuration of processing functions included in a diagnosis support device. [Diagram 5] FIG. 13 is a diagram illustrating an example of a data structure of a learning dataset. [Figure 6] FIG. 13 is a diagram for explaining a generation process of a lesion identification model. [Figure 7] 11 is a diagram for explaining a lesion detection process using a lesion identification model. FIG. [Figure 8] 13 is an example of a weighting table showing the correspondence between image feature amounts and weighting coefficients. [Figure 9] 13 is an example of a flowchart illustrating a procedure for a model generation process. [Figure 10] 13 is an example of a flowchart illustrating a procedure for a lesion detection process. [Figure 11] FIG. 13 is a diagram showing an example of a screen display of a detection result of a lesion area. [Figure 12] 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 13] FIG. 13 is a diagram for explaining a generation process of a lesion identification model. [Figure 14] 13 is an example of a weighting table showing the correspondence between image feature amounts and weighting coefficients. [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 illustrating a procedure for a lesion detection process. [Figure 17] FIG. 13 is a diagram for explaining a model generation process of a diagnosis support device according to a fourth embodiment. [Figure 18] 13 is an example of a weighting table showing the correspondence between image feature amounts and weighting coefficients. [Figure 19] 13 is an example of a flowchart illustrating a procedure for a model generation process. [Figure 20] 13 is an example of a flowchart illustrating a procedure for a lesion detection process. [Figure 21] 13 is an example of a weighting table showing the correspondence between image feature amounts and weighting coefficients. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. First Embodiment FIG. 1 is a diagram showing a configuration example and a processing example of an information processing device according to a first embodiment. The information processing device 1 shown in FIG. 1 detects a specific lesion area from a tomographic image of the inside of a human body. In the following description, as an example, the tomographic image is assumed to be an image captured at a certain interval in the height direction of the human body. To detect the lesion area, a lesion identification model is used that identifies whether or not each unit image area included in the input tomographic image is a lesion area. The unit image area is, for example, a pixel on a tomographic image, or a set of a predetermined number of adjacent pixels (image block) that is two or more. The information processing device 1 is also capable of generating such a lesion identification model by machine learning.

[0012] This information processing device 1 includes a storage unit 2 and a processing unit 3. The storage unit 2 is a storage area secured in a storage device (not shown) included in the information processing device 1. The storage unit 2 stores, for example, model parameters indicating a lesion identification model. The processing unit 3 is, for example, a processor. In this case, the following processing of the processing unit 3 is realized, for example, by the processor executing a predetermined program. The processing unit 3 executes a learning process for generating a lesion identification model by machine learning, and a lesion detection process for detecting a lesion area using the generated lesion identification model.

[0013] In the learning process, the following process is executed. The processing unit 3 classifies a plurality of first tomographic images obtained by photographing the inside of a plurality of first human bodies into a plurality of tomographic image groups having different medical findings. The processing unit 3 generates a plurality of lesion identification models that identify whether each unit image area included in the tomographic image to be identified is a lesion area or not by machine learning using different ones of the plurality of tomographic image groups as learning data.

[0014] In FIG. 1, as an example, a tomographic image group 11 including the above-mentioned multiple first tomographic images is prepared, and the first tomographic images included in this tomographic image group 11 are classified into two tomographic image groups 12a and 12b having different medical findings. For example, a finding related to an abnormal event that may change the pixel value relationship between a lesion area and a normal area in a tomographic image is used as the medical finding. In FIG. 1, as an example, the first tomographic images included in the tomographic image group 11 are classified into a tomographic image group 12a of normal findings in which no abnormal event has occurred, and a tomographic image group 12b of abnormal findings in which an abnormal event has occurred. In practice, the medical findings may be determined for each of multiple first tomographic images obtained by one shooting of one person.

[0015] Processing unit 3 generates lesion identification model 21 by machine learning using the tomographic images included in tomographic image group 12a as learning data. Processing unit 3 also generates lesion identification model 22 by machine learning using the tomographic images included in tomographic image group 12b as learning data.

[0016] Here, as an example of the abnormal event, fat accumulation in an organ such as the liver can be applied. In a CT image, a tumor in an organ region appears darker than a normal organ region (normal region). On the other hand, since fat regions also appear dark, when fat accumulates in an organ, the entire CT image becomes dark, and the difference in brightness between the lesion region (tumor region) and the normal region becomes small. In some cases, the brightness between the lesion region and the normal region is reversed. In this way, the relationship of pixel values ​​between the lesion region and the normal region in a CT image may change depending on whether fat is accumulated or not.

[0017] To address such problems, the above-mentioned processing by the processing unit 3 generates a lesion identification model 21 suitable for a tomographic image of a normal finding and a lesion identification model 22 suitable for a tomographic image of an abnormal finding. That is, the lesion identification model 21 can detect a lesion area from a tomographic image of a normal finding with higher accuracy than the lesion identification model generated using a CT image group containing a mixture of CT images of normal findings with no fat accumulation and CT images of abnormal findings with fat accumulation. On the other hand, the lesion identification model 22 can detect a lesion area from a tomographic image of an abnormal finding with higher accuracy than the lesion identification model generated using a CT image group containing a mixture of CT images of normal findings and CT images of abnormal findings.

[0018] After the lesion identification models 21, 22 are generated as described above, the following process is executed in the lesion processing. The processing unit 3 acquires a tomographic image group 13 including a plurality of second tomographic images obtained by photographing the inside of the second human body. The processing unit 3 calculates an image feature amount based on the acquired tomographic image group 13. When the abnormal event as described above is applied, for example, the average luminance in the organ region of each of the second tomographic images included in the tomographic image group 13 can be used as the image feature amount.

[0019] Processing unit 3 also inputs the second tomographic images included in tomographic image group 13 to each of the generated lesion identification models 21 and 22. This causes each of lesion identification models 21 and 22 to execute a lesion area identification process. Processing unit 3 obtains from lesion identification model 21 the probability that each unit image area included in the second tomographic images is a lesion area. Processing unit 3 also obtains from lesion identification model 22 the probability that each unit image area included in the second tomographic images is a lesion area.

[0020] Processing unit 3 calculates an integrated value by integrating the probabilities acquired from each of lesion identification models 21 and 22 for each unit image region included in the second tomographic images based on the calculated image feature amount. Processing unit 3 detects a lesion region from each of the second tomographic images based on the calculated integrated value. For example, if the integrated value is equal to or greater than a predetermined threshold, the corresponding unit image region is identified as a lesion region, and if the integrated value is less than the threshold, the corresponding unit image region is identified as a non-lesion region.

[0021] In the above process, for example, the occurrence degree of the abnormal event (degree of fat accumulation) in the plurality of second tomographic images is estimated based on the image feature amount. Then, in calculating the integrated value, processing unit 3 can perform weighted addition of the probabilities by assigning a larger weight to the probability from an appropriate lesion identification model according to the occurrence degree of the abnormal event. For example, when the image feature amount (average brightness) is equal to or greater than a predetermined threshold, it is estimated to be a normal finding, and therefore lesion identification model 22 assigns a larger weight to the probability from lesion identification model 21. On the other hand, when the image feature amount (average brightness) is less than the threshold, it is estimated to be an abnormal finding, and therefore lesion identification model 21 assigns a larger weight to the probability from lesion identification model 22.

[0022] By such processing, the processing unit 1 can accurately detect the lesion area from each of the second tomographic images regardless of the findings related to the abnormal events (i.e., regardless of the presence or absence of fat accumulation) for the plurality of second tomographic images inputted. Therefore, the detection accuracy of the lesion area from the tomographic images can be improved.

[0023] Second Embodiment Next, we will explain a system that can detect liver lesions from CT images. 2 is a diagram showing an example of the configuration of a diagnosis support system according to the second embodiment. The diagnosis support system shown in Fig. 2 is a system for supporting image diagnosis work by CT imaging, and includes a CT device 50 and a diagnosis support device 100.

[0024] 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 abdominal region including the liver while changing a position (slice position) in the height direction of the human body (direction perpendicular to the axial plane) at a predetermined interval.

[0025] The diagnosis support device 100 extracts a liver region from each tomographic image captured by the CT device 50, and detects a lesion region based on image information of the extracted region. In this embodiment, a tumor (e.g., hemangioma) in the liver is detected as the lesion region. The diagnosis support device 100, for example, displays information indicating the detection result of the lesion region on a display device. In this way, the diagnosis support device 100 supports the image diagnosis work of a user (e.g., a radiology doctor).

[0026] The diagnosis support device 100 detects a lesion area (or identifies a lesion) using a trained model generated by machine learning. The diagnosis support device 100 is also capable of executing a process of generating the trained model by machine learning.

[0027] 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.

[0028] 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.

[0029] 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.

[0030] 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.

[0031] 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.

[0032] 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.

[0033] 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.

[0034] 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.

[0035] As described above, the process of detecting a lesion area from a medical image can be performed using, for example, a trained model generated by machine learning. Fig. 3 is a diagram showing an example of a machine learning model for detecting a lesion area. Lesion identification model 60 shown in Fig. 3 is a trained model that, when a tomographic image 62 is input, identifies whether or not each pixel of tomographic image 62 is a specific lesion area. Here, it is assumed that tomographic image 62 is a tomographic image of an axial plane. In practice, lesion identification model 60 calculates the probability that each pixel of input tomographic image 62 is a lesion area, and identifies pixels with a probability equal to or greater than a predetermined value as a lesion area.

[0036] This lesion identification model 60 is generated by machine learning (e.g., deep learning) using the axial plane tomographic images 61a, 61b, 61c,... as training data. Labels indicating whether or not a pixel is a lesion area are added to each of the tomographic images 61a, 61b, 61c,..., and these labels are used as correct answer data during machine learning.

[0037] The diagnosis support device 100 of the present embodiment can detect a lesion area using such a lesion identification model 60. Another example of the lesion identification model is one that identifies whether or not a lesion area is present using a partial image (patch image) of a certain size as a unit. In this case, each tomographic image used as learning data is divided into patch images of a certain size, and a label indicating whether or not a lesion area is present is added to each divided patch image. Then, machine learning is performed using each patch image as teacher data and the label as correct answer data to generate a lesion identification model. During inference, a tomographic image input to the lesion identification model is divided into patch images of a certain size, and the lesion identification model identifies whether or not each patch image is a lesion area.

[0038] In lesion identification processing using the above-mentioned lesion identification model, it is considered that the greater the brightness difference between the lesion area and the normal area in the organ area on the CT image, the higher the accuracy of identifying the lesion area. In actual medical practice, CT images are taken using a contrast agent to emphasize the brightness difference between the lesion area and the normal area. However, the use of a contrast agent may cause side effects in the subject. For this reason, subjects who can use a contrast agent are limited to those in a relatively good condition, such as those who have no subjective symptoms due to the disease or those who are suspected of having a disease.

[0039] For these reasons, it is desirable to be able to accurately identify diseased areas from CT images taken without using contrast agents (here referred to as "plain CT images"). Furthermore, plain CT images are often taken at a stage when no specific disease is suspected, such as during health checkups. Being able to detect serious diseases such as cancer from CT images taken at such a stage would lead to the true early detection of disease. Furthermore, if diseased areas could be identified automatically and accurately from plain CT images, it would become possible to detect diseases even in medical institutions that do not have imaging diagnostic specialists on-site.

[0040] However, in plain CT images, the difference in brightness between the lesion area and the normal area in the organ area is small, and the lesion area may not be identified with high accuracy. For example, when detecting a tumor in the liver, the tumor area appears darker in plain CT images than the normal area in the liver because the brightness value of the tumor area is lower. However, if fat is accumulated in the liver, the fat also appears dark on the plain CT image, and the brightness value of the entire liver area decreases. As a result, the difference in brightness between the tumor area and the normal area becomes small, making it difficult to identify the tumor area with high accuracy. In some cases, the relationship of brightness between the tumor area and the normal area is reversed, making it even more difficult to identify the tumor area with high accuracy.

[0041] Considering these points, it is considered difficult to generate a lesion classification model with high classification accuracy when the plain CT images used as learning data during machine learning contain a mixture of plain CT images that contain fat and those that do not. This is because machine learning is performed using plain CT images with variations in the brightness difference between lesion areas and normal areas, or plain CT images with different brightness magnitude relationships between lesion areas and normal areas.

[0042] Therefore, in this embodiment, the CT image group of the plain CT images used as learning data is classified into a plurality of CT image groups each having a different occurrence rate of abnormal events that may change the brightness relationship between the lesion area and the normal area. In the above example, the accumulation of fat is the abnormal event. Such classification of the CT image group may be performed based on the findings of the radiologist.

[0043] The diagnosis support device 100 according to this embodiment generates a plurality of lesion identification models by using, as learning data, each of the CT image groups different from each other among the plurality of CT image groups classified in this manner. In this way, a plurality of lesion identification models are generated according to differences in the features of the input CT images.

[0044] During inference, the diagnosis support device 100 acquires a group of CT images using simple CT images, and estimates the occurrence rate of abnormal events in the acquired group of CT images from the image features of the group of CT images. At the same time, the diagnosis support device 100 executes a lesion identification process for the group of CT images using each of the above-mentioned multiple lesion identification models, and acquires a probability value for each pixel of a CT image included in the group of CT images that is a lesion area from each lesion identification model. The diagnosis support device 100 integrates the acquired probability values ​​at a ratio according to the image features to output a final identification result.

[0045] In probability value integration, the final classification result can be determined by increasing the ratio of probability values ​​from lesion classification models that are suitable for the occurrence of abnormal events in a group of CT images based on image features. The final classification result can also be determined by selecting a probability value from one lesion classification model that is optimal for the occurrence of abnormal events in a group of CT images. This makes it possible to improve the accuracy of lesion classification from plain CT images.

[0046] The following explanation mainly illustrates the use of a lesion identification model generated using a CT image of an abnormal finding in which fat has accumulated, and a lesion identification model generated using a CT image of a normal finding in which fat has not accumulated.

[0047] 4 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 model parameter storage unit 120, learning processing units 131 and 132, an image acquisition unit 141, an organ region extraction unit 142, a lesion identification unit 143, an image feature amount calculation unit 144, and a detection result output unit 145.

[0048] The learning data storage unit 110 and the model parameter storage unit 120 are storage areas secured in a storage device included in the diagnosis support device 100, such as the RAM 102 or the HDD 103. The learning data storage unit 110 stores a plurality of learning data sets 111 used for generating a lesion identification model by machine learning. The learning data sets 111 include a set of CT images (tomographic images) for one subject. The learning data sets 111 include a data set including a set of CT images of normal findings, and a data set including a set of CT images of abnormal findings (findings indicating the presence of fat). The diagnosis support device 100 uses these learning data sets 111 to generate two lesion identification models according to differences in findings. Note that all of the CT images included in each learning data set 111 are simple CT images taken without using a contrast agent.

[0049] The model parameter storage unit 120 stores model parameters 121 corresponding to an organ region extraction model, which is a trained model for extracting an organ region (liver region), and model parameters 122a and 122b corresponding to the above two lesion identification models, respectively. The model parameters stored in the model parameter storage unit 120 are data for forming a trained model, and include, for example, weighting coefficients between nodes on a neural network.

[0050] The processing of the learning processing units 131, 132, image acquisition unit 141, organ area extraction unit 142, lesion identification unit 143, image feature calculation unit 144 and detection result output unit 145 is realized, for example, by the processor 101 executing an application program.

[0051] The learning processing units 131 and 132 operate in a learning phase in which an organ region extraction model and a lesion identification model are generated by machine learning. The learning processing unit 131 generates an organ region extraction model by performing machine learning using each learning data set 111 stored in the learning data storage unit 110. The learning processing unit 131 stores model parameters 121 indicating the generated organ region extraction model in the model parameter storage unit 120.

[0052] The learning processing unit 132 generates a lesion identification model by performing machine learning using each learning data set 111 stored in the learning data storage unit 110. At this time, the learning processing unit 132 generates a lesion identification model corresponding to a normal finding using the learning data set 111 including a CT image set of a normal finding (learning data set of a normal finding). At the same time, the learning processing unit 132 generates a lesion identification model corresponding to an abnormal finding using the learning data set 111 including a CT image set of an abnormal finding (learning data set of an abnormal finding). The learning processing unit 132 stores model parameters 122a and 122b corresponding to the two generated lesion identification models, respectively, in the model parameter storage unit 120.

[0053] Image acquisition unit 141, organ area extraction unit 142, lesion identification unit 143, image feature calculation unit 144 and detection result output unit 145 operate in an inference phase to detect a lesion area using the generated organ area extraction model and lesion identification model.

[0054] The image acquisition unit 141 acquires a CT image set obtained by imaging a diagnostic subject from the CT device 50. The acquired CT image set is an image set of plain CT images captured without using a contrast agent.

[0055] The organ region extraction unit 142 uses an organ region extraction model based on the model parameters 121 to extract a CT image including an organ region (liver region) from the set of CT images. Lesion identification unit 143 executes lesion identification processing for each CT image including an organ region using a lesion identification model based on model parameters 122a, and outputs a probability value for each pixel that the region is a lesion region. Also, lesion identification unit 143 executes lesion identification processing for each CT image including an organ region using a lesion identification model based on model parameters 122b, and outputs a probability value for each pixel that the region is a lesion region.

[0056] The image feature amount calculation unit 144 calculates an image feature amount from an organ region among image regions of the CT images included in the CT image set. In the present embodiment, the average luminance in the organ region is calculated as the image feature amount.

[0057] The detection result output unit 145 integrates the lesion identification results from the two lesion identification models and outputs the final lesion area detection result. Specifically, for each pixel of the CT image including the organ area, the detection result output unit 145 performs weighted addition on the probability values ​​calculated by the two lesion identification models using a weighting coefficient according to the calculation result of the image feature amount. The detection result output unit 145 compares the calculated value of the weighted addition with a predetermined threshold value, and identifies pixels whose calculated value is equal to or greater than the threshold value as pixels in the lesion area.

[0058] 5 is a diagram showing an example of the data configuration of the learning data set. The learning data set 111 stored in the learning data storage unit 110 includes a CT image set 112, a lesion area image set 113, an organ area image set 114, and finding information 115.

[0059] The CT image set 112 includes a plurality of CT images 112a (tomographic images) obtained by one imaging of a single imaging subject. Lesion area image set 113 includes lesion area images 113a corresponding to the multiple CT images 112a included in CT image set 112. Lesion area image 113a is an image in which a label indicating whether or not a pixel in corresponding CT image 112a is associated with the pixel. Each label in lesion area image 113a is used as a correct answer label when generating a lesion identification model by machine learning.

[0060] The organ region image set 114 includes organ region images 114a corresponding to the multiple CT images 112a included in the CT image set 112. The organ region image 114a is an image in which a label indicating whether or not it is an organ region (liver region) is associated with each pixel of the corresponding CT image 112a. Each label of the organ region image 114a is used as a correct answer label when generating an organ region extraction model by machine learning.

[0061] The finding information 115 is information indicating whether a finding is normal or abnormal for a plurality of CT images 112a included in the CT image set 112. The finding information 115 is generated for each CT image set 112. For example, when a radiologist comprehensively views the CT images 112a included in the CT image set 112, if it is determined that a certain amount of fat has accumulated in an organ, the finding is determined to be abnormal, and if not, the finding is determined to be normal. In this way, the finding information 115 is information generated based on the medical findings of the radiologist.

[0062] 6 is a diagram for explaining the generation process of the lesion identification model. The learning data set 111 stored in the learning data storage unit 110 is classified into a learning data set 111 including a CT image set 112 of normal findings (learning data set 111 of normal findings) and a learning data set 111 including a CT image set 112 of abnormal findings (learning data set 111 of abnormal findings) based on the finding information 115.

[0063] The learning processing unit 132 extracts a learning data set 111 of normal findings from the learning data set 111 stored in the learning data storage unit 110 based on the finding information 115. The learning processing unit 132 generates a lesion identification model A1 corresponding to a normal finding by performing machine learning using each CT image 112a and lesion area image 113a included in the extracted learning data set 111. In practice, the CT image 112a including an organ area among the CT images 112a included in the CT image set 112 of normal findings and the lesion area image 113a corresponding to this are used for the machine learning. As a result, model parameters 122a indicating the lesion identification model A1 are stored in the model parameter storage unit 120.

[0064] Furthermore, the learning processing unit 132 extracts a learning data set 111 of an abnormal finding from the learning data set 111 stored in the learning data storage unit 110 based on the finding information 115. The learning processing unit 132 generates a lesion identification model A2 corresponding to the abnormal finding by performing machine learning using each CT image 112a and lesion area image 113a included in the extracted learning data set 111. In practice, the CT image 112a including an organ area among the CT images 112a included in the CT image set 112 of the abnormal finding and the lesion area image 113a corresponding thereto are used for the machine learning. As a result, model parameters 122b indicating the lesion identification model A2 are stored in the model parameter storage unit 120.

[0065] Through the above model generation process, a lesion identification model A1 suitable for detecting lesions from a group of CT images in which fat is not accumulated in organ areas, and a lesion identification model A2 suitable for detecting lesions from a group of CT images in which fat is included in organ areas are individually generated.

[0066] FIG. 7 is a diagram for explaining the lesion detection process using the lesion identification model. A CT apparatus 50 captures an image of a subject to be diagnosed, and a CT image set 151 is obtained. From this CT image set 151, CT images including organ regions are extracted using an organ region extraction model, and each of the extracted CT images is input to both of the lesion identification models A1 and A2.

[0067] Lesion identification model A1 executes lesion identification processing on a pixel-by-pixel basis for each input CT image, and outputs a probability value indicating the probability that each pixel is a lesion area. As a result, for example, probability image set 161a including probability images corresponding to the input CT images is generated. The probability images included in probability image set 161a are images in which a probability value is associated with each pixel of the corresponding CT image.

[0068] In addition, lesion identification model A2 executes lesion identification processing on a pixel-by-pixel basis for each input CT image, and outputs a probability value indicating the probability that each pixel is a lesion area. As a result, for example, probability image set 161b including probability images corresponding to the input CT images is generated. The probability images included in probability image set 161b are images in which a probability value is associated with each pixel of the corresponding CT image.

[0069] For example, when the lesion identification models A1 and A2 are machine learning models using a neural network, the above probability values ​​are output from the final layer (output layer) of the neural network.

[0070] Furthermore, each CT image including an organ region is also input to the image feature amount calculation unit 144. The image feature amount calculation unit 144 extracts the luminance value of each pixel in the organ region from the image region of each input CT image, and calculates the average value (average luminance) of the extracted luminance values ​​as the image feature amount.

[0071] Detection result output unit 145 integrates the lesion identification results from lesion identification models A1 and A2, and outputs the final lesion area detection result. At this time, detection result output unit 145 integrates the probability values ​​in the probability images in probability image set 161a and the probability values ​​in the probability images in probability image set 161b for each pixel of the processed CT image at a ratio according to the calculated image feature amount (average luminance). In this integration process, weighted addition is performed on the probability values ​​of each probability image using a weighting coefficient according to the image feature amount.

[0072] The detection result output unit 145 compares the integrated probability value with a predetermined threshold value to output a final detection result indicating whether or not each pixel of the CT image is a lesion area. For example, a result image is generated in which information indicating whether or not each pixel of the CT image to be processed is associated with the pixel. The result image set 171 shown in FIG. 7 includes result images corresponding to the multiple CT images to be processed.

[0073] According to the above lesion detection process, the diagnosis support device 100 can estimate the degree of fat accumulation in the CT images based on the image feature values ​​calculated from the input CT images. For example, it is estimated that the lower the average brightness, the more fat has accumulated. Then, the diagnosis support device 100 can appropriately integrate the lesion identification results from the lesion identification models A1 and A2 according to the degree of fat accumulation estimated based on the image feature values ​​to obtain a final lesion identification result.

[0074] Therefore, it is possible to detect lesion areas from plain CT images with high accuracy. For example, even when the difference in brightness between lesion areas and normal areas in an organ area is small, the magnitude relationship of brightness between lesion areas and normal areas is not uniform, or the brightness in the lesion area is not uniform, the detection accuracy of the lesion area can be improved.

[0075] Fig. 8 is an example of a weighting table showing the correspondence between image features and weighting coefficients. Detection result output unit 145 can integrate the classification results of lesion classification models A1 and A2 by referring to weighting table 181 as shown in Fig. 8, for example.

[0076] In FIG. 8, as an example, the region where the calculated average luminance can be obtained is divided into three regions using threshold values Ma1 and Ma2 (where Ma1 > Ma2), and weight coefficients for each of the lesion discrimination models A1 and A2 are associated with each of the divided regions. When the average luminance is greater than or equal to the threshold value Ma1, Xa1 and Xa2 are applied as the weight coefficients corresponding to the lesion discrimination models A1 and A2, respectively. When the average luminance is less than the threshold value Ma1 and greater than or equal to the threshold value Ma2, Ya1 and Ya2 are applied as the weight coefficients corresponding to the lesion discrimination models A1 and A2, respectively. When the average luminance is less than the threshold value Ma2, Za1 and Za2 are applied as the weight coefficients corresponding to the lesion discrimination models A1 and A2, respectively.

[0077] Here, the weight coefficients Xa1, Xa2, Ya1, Ya2, Za1, and Za2 each take values between 0 and 1, and Xa1 + Xa2 = Ya1 + Ya2 = Za1 + Za2 = 1. When the weight coefficient corresponding to one of the lesion discrimination models A1 and A2 is 1, the weight coefficient corresponding to the other becomes 0. In this case, the probability value output from one of the lesion discrimination models is used as it is.

[0078] Also, it is estimated that the lower the average luminance, the more fat is accumulated. Therefore, the lower the average luminance, the higher the weight coefficient of the lesion discrimination model A2 corresponding to the abnormal finding is set. Specifically, Xa1 > Ya1 > Za1 and Xa2 < Ya2 < Za2. Also, Xa1 > Xa2 and Za1 < Za2. The magnitude relationship between Ya1 and Ya2 may be determined according to the set values of the threshold values Ma1 and Ma2.

[0079] By using such weight coefficients, the detection result output unit 145 can increase the weight of the probability value by the lesion detection model that is closer to the conditions of the CT image group with the conditions of the CT image used as the learning data among the lesion discrimination models A1 and A2, and obtain the final lesion discrimination result. As a result, the lesion area can be accurately detected from a simple CT image.

[0080] In addition to the average brightness of the organ region in the CT image group corresponding to one person as described above, the image feature can also be, for example, the ratio of high brightness pixels in the organ region whose brightness is equal to or higher than a predetermined threshold. The ratio of such high brightness pixels is calculated as the ratio of the number of high brightness pixels contained in the organ region to the total number of pixels in the organ region in the CT image group. When the ratio of such high brightness pixels is used as the image feature, the weight coefficient of the lesion identification model A2 corresponding to the abnormal finding can be set to a higher value as the ratio of high brightness pixels is lower.

[0081] Also, FIG. 6 shows an example in which two lesion identification models A1 and A2 are generated using two CT image sets with different findings. However, as another example, three or more lesion identification models may be generated using three or more CT image sets with different findings. For example, as a lesion identification model corresponding to an abnormal finding, a lesion identification model based on a CT image set with a mild degree of fat accumulation and a lesion identification model based on a CT image set with a severe degree of fat accumulation may be generated. In this case, a lesion identification model corresponding to a normal finding and two lesion identification models corresponding to abnormal findings are used at the time of lesion detection. When three or more lesion identification models are used in this way, at the time of lesion detection, the lesion identification results by each of the three or more lesion identification models may be integrated at a ratio according to the image feature amount.

[0082] In addition to the above-mentioned fat accumulation, calcification in the liver is also considered as an abnormal event that can change the brightness relationship between the lesion area and the normal area. Liver areas with advanced calcification appear bright on CT images. Therefore, when calcification is advanced, the brightness value of the entire liver area becomes high. As a result, the brightness difference between the lesion area and the normal area becomes small, making it difficult to identify the lesion area with high accuracy.

[0083] For this reason, the diagnosis support device 100 may generate a lesion identification model corresponding to a normal finding using a CT image set without calcification, and generate a lesion identification model corresponding to an abnormal finding using a CT image set with calcification. In this case, for example, the proportion of high-luminance pixels described above can be used as an image feature when identifying a lesion. That is, the lesion identification results (probability values) from each lesion identification model are integrated at a ratio according to the proportion of high-luminance pixels in the input CT image set. At this time, the weight coefficient corresponding to the lesion identification model corresponding to the abnormal finding may be set to a higher value as the proportion of high-luminance pixels increases.

[0084] In the above description, the multiple lesion identification models are generated by considering the occurrence rate of one type of abnormal event that can change the luminance relationship between the lesion area and the normal area. However, as another example, multiple lesion identification models may be generated by considering the occurrence rate of multiple types of abnormal events that can change the luminance relationship between the lesion area and the normal area.

[0085] For example, both fat accumulation and calcification may be considered as abnormal events. In this case, for example, a lesion identification model corresponding to normal findings is generated using a CT image set in which neither fat accumulation nor calcification is observed. In addition, as lesion identification models corresponding to abnormal findings, a lesion identification model based on a CT image set in which fat accumulation is observed and a lesion identification model based on a CT image set in which calcification is observed are generated. When detecting a lesion, the lesion identification results from each lesion identification model may be integrated at a ratio according to the image feature amount.

[0086] Next, the process of the diagnosis support device 100 according to the second embodiment will be described with reference to a flowchart. First, the procedure of the model generation process for generating the above-mentioned lesion identification models A1 and A2 will be described with reference to Fig. 9. Fig. 9 is an example of a flowchart showing the procedure of the model generation process.

[0087] [Step S11] The learning processing unit 131 performs machine learning using all the CT images 112a stored in the learning data storage unit 110 and the organ area images 114a corresponding to each CT image 112a, thereby generating an organ area extraction model. In this machine learning, information on each pixel of the organ area image 114a is used as correct answer data. The learning processing unit 131 stores model parameters 121 indicating the generated organ area extraction model in the model parameter storage unit 120.

[0088] In this embodiment, as an example, the organ region extraction model is generated using the same learning data as that used when generating the lesion identification models A1 and A2. However, the organ region extraction model may be an organ region extraction model that has been generated in advance using other learning data.

[0089] [Step S12] The learning processing unit 132 extracts a learning data set 111 of normal findings from the learning data set 111 stored in the learning data storage unit 110, based on the finding information 115. Furthermore, the learning processing unit 132 extracts a CT image 112a including an organ region from the extracted learning data set 111, based on the organ region image 114a.

[0090] Learning processing unit 132 performs machine learning using extracted CT images 112a and lesion area images 113a corresponding to these CT images 112a to generate a lesion identification model A1 corresponding to a normal finding. In this machine learning, information on each pixel of lesion area image 113a is used as correct answer data. Learning processing unit 131 stores model parameters 122a indicating the generated lesion identification model A1 in model parameter storage unit 120.

[0091] [Step S13] The learning processing unit 132 extracts a learning data set 111 of abnormal findings from the learning data set 111 stored in the learning data storage unit 110, based on the finding information 115. Furthermore, the learning processing unit 132 extracts a CT image 112a including an organ region from the extracted learning data set 111, based on the organ region image 114a.

[0092] Learning processing unit 132 performs machine learning using extracted CT images 112a and lesion area images 113a corresponding to these CT images 112a to generate a lesion identification model A2 corresponding to the abnormal findings. In this machine learning, information on each pixel of lesion area image 113a is used as correct answer data. Learning processing unit 131 stores model parameters 122b indicating the generated lesion identification model A2 in model parameter storage unit 120.

[0093] Next, a procedure of a lesion detection process for detecting a lesion area using the generated lesion identification models A1 and A2 will be described with reference to Fig. 10. Fig. 10 is an example of a flowchart showing the procedure of the lesion detection process.

[0094] [Step S21] The image acquisition unit 141 acquires from the CT device 50 a set of CT images obtained by imaging a diagnostic subject. [Step S22] The organ region extraction unit 142 inputs each CT image included in the CT image set to the organ region extraction model based on the model parameters 121. As a result, the organ region extraction unit 142 identifies whether or not each pixel of each CT image is an organ region. For example, an organ region image is generated in which information indicating whether or not each pixel of the CT image is an organ region is associated with the pixel.

[0095] The organ region extraction unit 142 extracts a CT image including an organ region (a CT image including one or more pixels of the organ region) from the CT image set. In the next step S23 and thereafter, processing is performed using the extracted CT image.

[0096] [Step S23] The image feature amount calculation unit 144 obtains the pixel values ​​of each pixel in the organ region from all CT images including the organ region, converts the obtained pixel values ​​into luminance values, and calculates the average value of the converted luminance values, i.e., the average luminance in the organ region.

[0097] [Step S24] The lesion identifying unit 143 selects a CT image to be processed from among CT images that include organ regions. [Step S25] Lesion identifier 143 inputs the CT image to be processed to lesion identification model A1 based on model parameters 122a, and then outputs a probability value indicating that each pixel in the CT image is a lesion area.

[0098] [Step S26] Lesion identifier 143 inputs the CT image to be processed to lesion identification model A2 based on model parameters 122b, and outputs a probability value indicating that each pixel in the CT image is a lesion area.

[0099] [Step S27] Detection result output unit 145 obtains a weighting factor corresponding to the average brightness calculated in step S23 from weight table 181. As a result, weighting factors corresponding to lesion identification models A1 and A2 are obtained. Detection result output unit 145 multiplies the probability values ​​output in steps S25 and S26 by the weighting factors corresponding to lesion identification models A1 and A2, respectively, and calculates an integrated probability value for each pixel by adding the multiplication results obtained. Detection result output unit 145 compares the calculated integrated probability value with a predetermined threshold, and identifies pixels whose integrated probability value is equal to or greater than the threshold as a lesion area, and identifies pixels whose integrated probability value is less than the threshold as a normal area. Detection result output unit 145 generates and outputs a result image in which information indicating whether or not each pixel of a CT image is a lesion area is associated with the pixel, for example.

[0100] [Step S28] The lesion identification unit 143 determines whether all CT images including organ regions have been selected. If there are unselected CT images, the process proceeds to step S24, where one unselected CT image is selected. On the other hand, if all the corresponding CT images have been selected, the lesion detection process ends.

[0101] 11 is a diagram showing an example of a screen display of the detection result of the lesion area. The detection result output unit 145 can output information indicating the detection result of the lesion area by using a result image generated for each CT image including an organ area. For example, the detection result output unit 145 can display a result display screen 70 as shown in FIG. 11 on the display device 104a. The result display screen 70 includes a slice selection unit 71, a CT image display unit 72, and a lesion detection area display unit 73.

[0102] In slice selection section 71, by moving handle 71a on a slider, a slice plane of a CT image to be displayed in CT image display section 72 can be selected. In CT image display section 72, a CT image corresponding to the slice plane selected in slice selection section 71 is displayed. In this CT image, a lesion area is displayed based on a result image corresponding to the same slice plane. In FIG. 11, as an example, a rectangle 72b surrounding lesion area 72a in the CT image is superimposed. As another display method, lesion area 72a may be displayed in a specific color. This lesion area 72a is a pixel area associated with information indicating the lesion area in the result image. Furthermore, organ area 74 may be displayed in another specific color.

[0103] Lesion detection area display unit 73 shows the range of slice planes on which lesion areas are detected (slice planes corresponding to resultant images containing one or more pixels of information indicating a lesion area) within the movable area of ​​handle 71a in slice selection unit 71. The display of lesion detection area display unit 73 allows the user to easily recognize slice planes on which lesion areas are detected, quickly display tomographic images corresponding to those slice planes, and confirm the lesion area on the tomographic images.

[0104] Third embodiment Next, an example in which the processing of the diagnosis support device 100 according to the second embodiment is partially modified will be described as a third embodiment. The third embodiment differs from the second embodiment in the following respects. In the third embodiment, a learning CT image set is classified into a plurality of CT image groups according to the range of image feature amounts calculated from the CT image set. Then, a plurality of lesion identification models are generated by using each of the classified CT image groups as learning data.

[0105] In the following description, it is assumed, as an example, that a training CT image set is classified into two CT image groups and two lesion classification models are generated. Fig. 12 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. 12, the same components as those in Fig. 4 are denoted by the same reference numerals.

[0106] The diagnosis support device 100a according to the third embodiment further includes an image feature amount calculation unit 133 in addition to the learning processing units 131 and 132, the image acquisition unit 141, the organ region extraction unit 142, the lesion identification unit 143, the image feature amount calculation unit 144, and the detection result output unit 145 shown in Fig. 4. The processing of the image feature amount calculation unit 133 is realized by a processor included in the diagnosis support device 100a executing a predetermined program. The image feature amount calculation unit 133 calculates the same type of image feature amount as the image feature amount calculation unit 144 from each of the CT image sets 112 stored in the learning data storage unit 110.

[0107] Moreover, CT image set 112 stored in learning data storage unit 110 is classified into a CT image group whose calculated image feature amount falls within a first range and a CT image group whose calculated image feature amount falls within a second range. Learning processing unit 132 generates a lesion identification model B1 (see FIG. 13) by machine learning using the CT images included in one CT image group as learning data, and stores model parameters 123a indicating the generated lesion identification model B1 in model parameter storage unit 120. Learning processing unit 132 also generates a lesion identification model B2 (see FIG. 13) by machine learning using the CT images included in the other CT image group as learning data, and stores model parameters 123b indicating the generated lesion identification model B2 in model parameter storage unit 120.

[0108] Lesion identification unit 143 calculates a probability value for each pixel that it is a lesion area using lesion identification models B1 and B2 based on model parameters 123a and 123b. Detection result output unit 145 integrates the calculated probability values ​​at a ratio according to the image features calculated by image feature calculation unit 144, and outputs a final lesion detection result.

[0109] FIG. 13 is a diagram for explaining the generation process of the lesion identification model. The image feature amount calculation unit 133 calculates the same type of image feature amount as the image feature amount calculation unit 144, from each of the CT image sets 112 stored in the learning data storage unit 110. Here, it is assumed that the average luminance in the organ region is calculated as the image feature amount.

[0110] The learning processing unit 132 classifies the CT image set 112 stored in the learning data storage unit 110 into a first CT image group whose image feature amount falls within a first range and a second CT image group whose image feature amount falls within a second range. The learning processing unit 132 generates a lesion identification model B1 by machine learning using the first CT image group as learning data, and generates a lesion identification model B2 by machine learning using the second CT image group as learning data.

[0111] The first range for the image feature amount is a range equal to or greater than a threshold value Nb1, and the second range is a range equal to or less than a threshold value Nb2. In this case, the first CT image group substantially corresponds to the CT image group of normal findings in the second embodiment, and the second CT image group substantially corresponds to the CT image group of abnormal findings in the second embodiment. Therefore, lesion identification model B1 substantially corresponds to normal findings, and lesion identification model B2 substantially corresponds to abnormal findings.

[0112] Here, the first range and the second range may be divided by one threshold. However, in this embodiment, the first and second ranges are set using two thresholds Nb1 and Nb2, and Nb2>Nb1. As a result, the thresholds Nb1 and Nb2 are set so that a part of the lower limit side of the first range overlaps with a part of the upper limit side of the second range.

[0113] The range of image feature values ​​(average brightness) that can be calculated from a CT image set of normal findings and the range of image feature values ​​(average brightness) that can be calculated from a CT image set of abnormal findings are not necessarily divided with a specific threshold as the boundary. When the image feature values ​​calculated from the CT image set are values ​​near the boundary of these ranges, it is not necessarily clear whether the CT image set is a normal finding or an abnormal finding. By overlapping a part of the first range and a part of the second range with Nb2>Nb1, it is possible to generate lesion identification models B1 and B2 taking such ambiguity into consideration. In other words, even if the feature values ​​of the CT image set input at the time of lesion detection are values ​​near the overlapping area of ​​the first and second ranges, it is possible to improve the lesion identification accuracy by using the lesion identification models B1 and B2.

[0114] When detecting a lesion area, lesion identification models B1 and B2 may be used in place of lesion identification models A1 and A2 in Fig. 7, respectively, and the final lesion detection result may be output in a similar procedure to that in Fig. 7. Furthermore, in the integration process of the probability values ​​from lesion identification model B1 and lesion identification model B2, weighting coefficients such as those shown in the following Fig. 14 may be used.

[0115] FIG. 14 is an example of a weight table showing the correspondence between image feature amounts and weight coefficients. The detection result output unit 145 can integrate the respective discrimination results of the lesion discrimination models B1 and B2 by referring to a weight table 182 as shown in FIG. 14, for example.

[0116] In FIG. 14, as an example, the region where the calculated average luminance can be obtained is divided into three regions using thresholds Mb1 and Mb2 (where Mb1 > Mb2), and weight coefficients for each of the lesion discrimination models B1 and B2 are associated with each of the divided regions. When the average luminance is equal to or greater than the threshold Mb1, Xb1 and Xb2 are applied as the weight coefficients corresponding to the lesion discrimination models B1 and B2, respectively. When the average luminance is less than the threshold Mb1 and equal to or greater than the threshold Mb2, Yb1 and Yb2 are applied as the weight coefficients corresponding to the lesion discrimination models B1 and B2, respectively. When the average luminance is less than the threshold Mb2, Zb1 and Zb2 are applied as the weight coefficients corresponding to the lesion discrimination models B1 and B2, respectively.

[0117] Here, the weight coefficients Xb1, Xb2, Yb1, Yb2, Zb1, and Zb2 each take values between 0 and 1, and Xb1 + Xb2 = Yb1 + Yb2 = Zb1 + Zb2 = 1. When the weight coefficient corresponding to one of the lesion discrimination models B1 and B2 is 1, the weight coefficient corresponding to the other is 0. In this case, the probability value output from one of the lesion discrimination models is used as it is.

[0118] Also, it is estimated that the lower the average luminance, the more fat is accumulated. For this reason, the lower the average luminance, the higher the weight coefficient of the lesion discrimination model B2 corresponding to the abnormal finding is set. Specifically, Xb1 > Yb1 > Zb1 and Xb2 < Yb2 < Zb2. Also, Xb1 > Xb2 and Zb1 < Zb2. The magnitude relationship between Yb1 and Yb2 may be determined according to the set values of the thresholds Mb1 and Mb2.

[0119] By using such weighting coefficients, detection result output unit 145 increases the weight of the probability value of the lesion detection model B1 or B2 whose CT image conditions used as learning data are closer to the conditions of the input CT image group, thereby making it possible to obtain a final lesion identification result. As a result, it becomes possible to detect lesion areas from plain CT images with high accuracy.

[0120] The image feature calculated by the image feature calculation units 133, 144 may be, for example, the proportion of high-luminance pixels in the organ region, the luminance of which is equal to or greater than a predetermined threshold, in addition to the average luminance of the organ region in the CT image group corresponding to one person as described above. When such a proportion of high-luminance pixels is used as the image feature, during learning, a lesion identification model B1 is generated using a CT image set in which the proportion of high-luminance pixels is equal to or greater than a threshold Nb1a. Also, a lesion identification model B2 is generated using a CT image set in which the proportion of high-luminance pixels is equal to or less than a threshold Nb2a (>Nb1a). During lesion detection, the weight coefficient of the lesion identification model B2 may be set to a higher value as the proportion of high-luminance pixels decreases.

[0121] 13 shows an example in which two lesion identification models B1 and B2 are generated using two CT image sets with different ranges of image features. However, as another example, three or more lesion identification models may be generated using three or more CT image sets with different ranges of image features. In this case, when detecting a lesion, it is sufficient that the lesion identification results from each of the three or more lesion identification models are integrated in a ratio according to the image features.

[0122] Also, multiple types of image features may be used. For example, the image feature calculation units 133 and 144 calculate the above-mentioned average luminance and the ratio of high-luminance pixels. During learning, for example, Nbx is used as the threshold of the average luminance, and Nby is used as the threshold of the ratio of high-luminance pixels. In this case, the learning CT image set 112 is classified into, for example, a CT image group with average luminance ≥ Nbx and ratio of high-luminance pixels ≥ Nby, a CT image group with average luminance ≥ Nbx and ratio of high-luminance pixels < Nby, a CT image group with average luminance < Nbx and ratio of high-luminance pixels ≥ Nby, and a CT image group with average luminance < Nbx and ratio of high-luminance pixels < Nby. Then, individual lesion discrimination models are generated using the classified CT image groups respectively. During lesion detection, the lesion discrimination results by each lesion discrimination model may be integrated at a ratio corresponding to the combination of the average luminance and the ratio of high-luminance pixels based on the input CT image set.

[0123] Also, similar to the second embodiment, liver calcification may be considered as an abnormal event that can change the luminance relationship between the lesion area and the normal area. In this case, for example, the image feature calculation units 133 and 144 calculate the ratio of the above-mentioned high-luminance pixels. During learning, a lesion discrimination model B1 is generated using a CT image set with the ratio of high-luminance pixels being equal to or less than the threshold Nb2b. Also, a lesion discrimination model B2 is generated using a CT image set with the ratio of high-luminance pixels being equal to or greater than the threshold Nb1b (<Nb2a). During lesion detection, the higher the ratio of high-luminance pixels, the higher the weight coefficient of the lesion discrimination model B2 may be set.

[0124] Next, the processing of the diagnostic support device 100a according to the third embodiment will be described using a flowchart. First, the procedure of the model generation process for generating the above-mentioned lesion discrimination models B1 and B2 will be described with reference to FIG. 15. FIG. 15 is an example of a flowchart showing the procedure of the model generation process.

[0125] [Step S31] The learning processing unit 131 performs machine learning using all the CT images 112a stored in the learning data storage unit 110 and the organ area images 114a corresponding to each CT image 112a, to generate an organ area extraction model. In this machine learning, information on each pixel of the organ area image 114a is used as correct answer data. The learning processing unit 131 stores model parameters 121 indicating the generated organ area extraction model in the model parameter storage unit 120.

[0126] [Step S32] The image feature amount calculation unit 133 calculates the average luminance in the organ region for each of the CT image sets 112 stored in the training data storage unit 110. [Step S33] The learning processing unit 132 extracts CT image sets 112 whose calculated average brightness is equal to or greater than a threshold value Nb1 from the CT image sets 112 stored in the learning data storage unit 110. The learning processing unit 132 extracts CT images 112a including an organ region from the extracted CT image sets 112 based on the organ region image 114a.

[0127] Learning processing unit 132 performs machine learning using extracted CT images 112a and lesion area images 113a corresponding to these CT images 112a to generate a lesion identification model B1 corresponding to normal findings. In this machine learning, information on each pixel of lesion area image 113a is used as correct answer data. Learning processing unit 131 stores model parameters 123a indicating the generated lesion identification model B1 in model parameter storage unit 120.

[0128] [Step S34] The learning processing unit 132 extracts CT image sets 112 whose calculated average brightness is equal to or less than a threshold value Nb2 from the CT image sets 112 stored in the learning data storage unit 110. The learning processing unit 132 extracts CT images 112a including an organ region from the extracted CT image sets 112 based on the organ region image 114a.

[0129] Learning processing unit 132 performs machine learning using extracted CT images 112a and lesion area images 113a corresponding to these CT images 112a to generate a lesion identification model B2 corresponding to the abnormal findings. In this machine learning, information on each pixel of lesion area image 113a is used as correct answer data. Learning processing unit 131 stores model parameters 123b indicating the generated lesion identification model B2 in model parameter storage unit 120.

[0130] Next, a procedure of a lesion detection process for detecting a lesion area using the generated lesion identification models B1 and B2 will be described with reference to Fig. 16. Fig. 16 is an example of a flowchart showing the procedure of the lesion detection process.

[0131] In Fig. 16, the same step numbers are assigned to the processing steps having the same processing contents as those in Fig. 10. In Fig. 16, steps S25a to S27a are executed instead of steps S25 to S27 in Fig. 10.

[0132] [Step S25a] Lesion identifier 143 inputs a CT image to be processed to lesion identifier model B1 based on model parameters 123a, and outputs a probability value that each pixel in the CT image is a lesion area.

[0133] [Step S26a] Lesion identifier 143 inputs a CT image to be processed to lesion identifier model B2 based on model parameters 123b, and outputs a probability value indicating that each pixel in the CT image is a lesion area.

[0134] [Step S27a] Detection result output unit 145 obtains a weighting factor corresponding to the average brightness calculated in step S23 from weight table 182. As a result, weighting factors corresponding to lesion identification models B1 and B2 are obtained. Detection result output unit 145 multiplies the probability values ​​output in steps S25a and S26a by the weighting factors corresponding to lesion identification models B1 and B2, respectively, and calculates an integrated probability value for each pixel by adding the multiplication results obtained. Detection result output unit 145 compares the calculated integrated probability value with a predetermined threshold, and identifies pixels whose integrated probability value is equal to or greater than the threshold as a lesion area, and identifies pixels whose integrated probability value is less than the threshold as a normal area. Detection result output unit 145 generates and outputs a result image in which information indicating whether or not each pixel of a CT image is a lesion area is associated with the pixel, for example.

[0135] [Fourth embodiment] The lesion detection process may be performed using all of the lesion identification models A1 and A2 generated in the second embodiment and the lesion identification models B1 and B2 generated in the third embodiment. Furthermore, lesion identification models C1 and C2 (see FIG. 17) generated in consideration of both differences in findings and differences in image features may be used. In the following fourth embodiment, an example of a diagnosis support device in which the lesion detection process is performed using these lesion identification models A1, A2, B1, B2, C1, and C2 will be described.

[0136] Fig. 17 is a diagram for explaining a model generation process of a diagnosis support device according to the fourth embodiment. A diagnosis support device 100b according to the fourth embodiment includes learning processing units 131 and 132, an image feature amount calculation unit 133, an image acquisition unit 141, an organ region extraction unit 142, a lesion identification unit 143, an image feature amount calculation unit 144, and a detection result output unit 145 shown in Fig. 12.

[0137] The learning processing unit 132 classifies the CT image set 112 included in the learning dataset 111 into a CT image set 112 of normal findings and a CT image set 112 of abnormal findings by a procedure similar to that of the second embodiment. The learning processing unit 132 generates a lesion classification model A1 by machine learning using the CT image set 112 of normal findings, and generates a lesion classification model A2 by machine learning using the CT image set 112 of abnormal findings.

[0138] Furthermore, the learning processing unit 132 classifies the CT image set 112 included in the learning data set 111 into a first CT image group whose image feature amount is equal to or greater than a threshold Nb1 and a second CT image group whose image feature amount is equal to or less than a threshold Nb2 (>Nb1) using a procedure similar to that of the third embodiment. The learning processing unit 132 generates a lesion identification model B1 by machine learning using the first CT image group, and generates a lesion identification model B2 by machine learning using the second CT image group.

[0139] Furthermore, the learning processing unit 132 extracts a third CT image group of CT image sets 112 whose image feature amount is equal to or greater than a threshold value Nc1 from the CT image set 112 of normal findings. Also, the learning processing unit 132 extracts a fourth CT image group of CT image sets 112 whose image feature amount is equal to or less than a threshold value Nc2 from the CT image set 112 of abnormal findings, where Nc2>Nc1. The learning processing unit 132 generates a lesion identification model C1 by machine learning using the third CT image group, and generates a lesion identification model C2 by machine learning using the fourth CT image group.

[0140] Lesion identification unit 143 calculates a probability value for each pixel that it is a lesion area using lesion identification models A1, A2, B1, B2, C1, and C2. Detection result output unit 145 integrates the calculated probability values ​​at a ratio according to the image features calculated by image feature calculation unit 144, and outputs a final lesion detection result.

[0141] Here, in real CT images, special cases may occur in which the relationship between findings and image features does not match the theoretical relationship, such as when fat is accumulated but the organ area is relatively bright. The lesion identification models C1 and C2 are trained models for enabling accurate detection of lesion areas even when such special cases occur. For example, by setting the threshold value Nc1 to a relatively low value, the lesion identification model C1 can accurately identify lesion areas even when a CT image set is input in which the finding indicates that fat is not accumulated but the average brightness is relatively low. In addition, by setting the threshold value Nc2 to a relatively high value, the lesion identification model C2 can accurately identify lesion areas even when a CT image set is input in which the finding indicates that fat is accumulated but the average brightness is relatively high.

[0142] Fig. 18 is an example of a weighting table showing the correspondence between image features and weighting coefficients. Detection result output unit 145 can integrate the classification results of lesion classification models A1, A2, B1, B2, C1, and C2 by referring to weighting table 183 as shown in Fig. 18, for example.

[0143] 18, for example, the area where the calculated average brightness can be is divided into three areas using thresholds Mc1 and Mc2 (where Mc1>Mc2). Each divided area is associated with a weighting factor for each of the lesion identification models A1, A2, B1, B2, C1, and C2.

[0144] When the average brightness is equal to or greater than the threshold Mc1, Xc1 to Xc6 are applied as weighting factors corresponding to the lesion identification models A1, A2, B1, B2, C1, and C2, respectively. When the average brightness is less than the threshold Mc1 and equal to or greater than the threshold Mc2, Yc1 to Yc6 are applied as weighting factors corresponding to the lesion identification models A1, A2, B1, B2, C1, and C2, respectively. When the average brightness is less than the threshold Mc2, weighting factors Zc1 to Zc6 are applied as weighting factors corresponding to the lesion identification models A1, A2, B1, B2, C1, and C2, respectively.

[0145] Here, the weighting factors Xc1 to Xc6, Yc1 to Yc6, and Zc1 to Zc6 each take values between 0 and 1. And Xc1 + Xc2 + Xc3 + Xc4 + Xc5 + Xc6 = Yc1 + Yc2 + Yc3 + Yc4 + Yc5 + Yc6 = Zc1 + Zc2 + Zc3 + Zc4 + Zc5 + Zc6 = 1.

[0146] Furthermore, the lower the average luminance, the higher the weighting factor of the lesion discrimination model A2 among the lesion discrimination models A1 and A2 is set, and the higher the weighting factor of the lesion discrimination model B2 among the lesion discrimination models B1 and B2 is set. Specifically, Xc1 > Yc1 > Zc1, Xc2 < Yc2 < Zc2, Xc3 > Yc3 > Zc3, Xc4 < Yc4 < Zc4. Also, Xc1 > Xc2, Zc1 < Zc2, Xc3 > Xc4, Zc3 < Zc4. The magnitude relationship between Yc1 and Yc2, and the magnitude relationship between Yc3 and Yc4 may be determined according to the set values of the threshold values Mc1 and Mc2.

[0147] On the other hand, the lesion discrimination models C1 and C2 are generated by machine learning that excludes learning data (outlier learning data) of special cases such as normal but dark or bright with fat accumulation. By such machine learning, stronger lesion discrimination models for normal findings and abnormal findings are generated. Regarding the weighting factors, Xc5 > Yc5 > Zc5, Xc6 < Yc6 < Zc6, Xc5 > Xc6, Zc5 < Zc6. The magnitude relationship between Yc5 and Yc6 may be determined according to the set values of the threshold values Mc1 and Mc2.

[0148] By using such weighting factors, the detection result output unit 145 can increase the weight of the probability value by the lesion detection model that is close to the condition of the CT image group in which the condition of the CT image used as learning data is input among the lesion discrimination models A1, A2, B1, B2, C1, and C2, and obtain the final lesion discrimination result. As a result, the lesion area can be accurately detected from a simple CT image.

[0149] Next, the processing of the diagnostic support device 100b according to the fourth embodiment will be described using a flowchart. First, the procedure of the model generation process for generating the above lesion identification models A1, A2, B1, B2, C1, and C2 will be described with reference to Fig. 19. Fig. 19 is an example of a flowchart showing the procedure of the model generation process.

[0150] [Step S41] The learning processing unit 131 generates an organ region extraction model in the same procedure as step S11 in FIG. 9 and step S31 in FIG. 15, and stores model parameters 121 indicating the generated organ region extraction model in the model parameter storage unit 120.

[0151] [Step S42] The learning processing unit 131 generates a lesion identification model A1 in the same procedure as in step S12 of FIG.

[0152] [Step S43] The learning processing unit 131 generates a lesion identification model A2 in the same manner as in step S13 of FIG.

[0153] [Step S44] The image feature amount calculation unit 133 calculates the average brightness in the organ region for each of the CT image sets 112 stored in the training data storage unit 110 in the same manner as in step S32 in FIG.

[0154] [Step S45] The learning processing unit 132 generates a lesion identification model B1 in the same manner as in step S33 of FIG.

[0155] [Step S46] The learning processing unit 132 generates a lesion identification model B2 in the same manner as in step S34 of FIG.

[0156] [Step S47] From among the CT images 112a of normal findings used as learning data in step S42, the learning processor 132 extracts CT images 112a whose average brightness calculated from the corresponding CT image set in step S44 is equal to or greater than the threshold value Nc1. The learning processor 132 generates a lesion identification model C1 by performing machine learning using the extracted CT images 112a and lesion area images 113a corresponding to these CT images 112a. In this machine learning, information on each pixel of the lesion area images 113a is used as correct answer data. The learning processor 131 stores model parameters indicating the generated lesion identification model C1 in the model parameter storage unit 120.

[0157] [Step S48] The learning processing unit 132 extracts CT images 112a having an average brightness calculated from the corresponding CT image set in step S44 that is equal to or less than a threshold value Nc2 from among the CT images 112a of abnormal findings used as learning data in step S43. The learning processing unit 132 generates a lesion identification model C2 by performing machine learning using the extracted CT images 112a and lesion area images 113a corresponding to these CT images 112a. In this machine learning, information on each pixel of the lesion area images 113a is used as correct answer data. The learning processing unit 131 stores model parameters indicating the generated lesion identification model C2 in the model parameter storage unit 120.

[0158] Next, a procedure of a lesion detection process for detecting a lesion area using the generated lesion identification models A1, A2, B1, B2, C1, and C2 will be described with reference to Fig. 20. Fig. 20 is an example of a flowchart showing the procedure of the lesion detection process.

[0159] In the process of Fig. 20, first, the processes of steps S21 to S23 in Fig. 10 are executed. As a result, the average brightness in the organ region of the CT image set is calculated. After that, the processes of step S51 and after are executed.

[0160] [Step S51] The lesion identifying unit 143 selects a CT image to be processed from among CT images that include organ regions. [Step S52] Lesion identifier 143 inputs a CT image to be processed to lesion identification model A1 based on model parameters 122a, and outputs a probability value indicating that each pixel in the CT image is a lesion area.

[0161] [Step S53] Lesion identifier 143 inputs the CT image to be processed to lesion identification model A2 based on model parameters 122b, and outputs a probability value indicating that each pixel in the CT image is a lesion area.

[0162] [Step S54] Lesion identifier 143 inputs the CT image to be processed to lesion identification model B1 based on model parameters 123a, and then outputs a probability value indicating that each pixel in the CT image is a lesion area.

[0163] [Step S55] Lesion identifier 143 inputs the CT image to be processed to lesion identification model B2 based on model parameters 123b, and then outputs a probability value indicating that each pixel in the CT image is a lesion area.

[0164] [Step S56] The lesion identifying unit 143 inputs the CT image to be processed to the lesion identification model C1, and outputs the probability value that each pixel in the CT image is a lesion area. [Step S57] Lesion identifier 143 inputs the CT image to be processed to lesion identification model C2, and outputs a probability value indicating that each pixel in the CT image is a lesion area.

[0165] [Step S58] Detection result output unit 145 obtains a weighting factor corresponding to the average brightness calculated in step S23 from weight table 183. As a result, weighting factors corresponding to lesion identification models A1, A2, B1, B2, C1, and C2 are obtained. Detection result output unit 145 multiplies the probability values ​​output in steps S52 to S57 by the weighting factors corresponding to lesion identification models A1, A2, B1, B2, C1, and C2, respectively, and calculates an integrated probability value for each pixel by adding up the multiplication results obtained. Detection result output unit 145 compares the calculated integrated probability value with a predetermined threshold, and identifies pixels whose integrated probability value is equal to or greater than the threshold as a lesion area, and identifies pixels whose integrated probability value is less than the threshold as a normal area. Detection result output unit 145 generates and outputs a result image in which, for example, information indicating whether or not each pixel of the CT image is a lesion area is associated with the pixel.

[0166] [Step S59] The lesion identification unit 143 determines whether all CT images including organ regions have been selected. If there are any unselected CT images, the process proceeds to step S51, where one unselected CT image is selected. On the other hand, if all the corresponding CT images have been selected, the lesion detection process ends.

[0167] The diagnosis support device may execute the lesion detection process by combining two pairs out of the pair of lesion identification models A1, A2, the pair of lesion identification models B1, B2, and the pair of lesion identification models C1, C2, for example.

[0168] For example, lesion detection processing is performed using lesion identification models A1, A2 and lesion identification models C1, C2. It can be said that lesion identification models C1, C2 cover special cases in which lesion identification models A1, A2 cannot accurately identify lesion areas. Therefore, by using lesion identification models A1, A2 and lesion identification models C1, C2, it is possible to compensate for the shortcomings of lesion identification models A1, A2 and to improve the accuracy of identifying lesion areas.

[0169] Furthermore, lesion detection processing may be performed using lesion identification models B1, B2 and lesion identification models C1, C2. It can be said that lesion identification models C1, C2 cover special cases in which lesion identification models B1, B2 cannot accurately identify lesion areas. For this reason, by using lesion identification models B1, B2 and lesion identification models C1, C2, it is possible to compensate for the shortcomings of lesion identification models B1, B2 and to improve the accuracy of identifying lesion areas.

[0170] As an example, FIG. 21 shows weighting coefficients when lesion identification models A1, A2 and lesion identification models C1, C2 are used. Fig. 21 is an example of a weighting table showing the correspondence between image features and weighting coefficients. Detection result output unit 145 can integrate the classification results of lesion classification models A1, A2, C1, and C2 by referring to weighting table 184 as shown in Fig. 21, for example.

[0171] 21, for example, the area where the calculated average brightness can be is divided into three areas using thresholds Md1 and Md2 (where Md1>Md2). Each divided area is associated with a weighting factor for each of the lesion identification models A1, A2, C1, and C2.

[0172] When the average brightness is equal to or greater than threshold value Md1, weighting factors Xd1 to Xd4 are applied as weighting factors corresponding to lesion identification models A1, A2, C1, and C2, respectively. When the average brightness is less than threshold value Md1 and equal to or greater than threshold value Md2, weighting factors Yd1 to Yd4 are applied as weighting factors corresponding to lesion identification models A1, A2, C1, and C2, respectively. When the average brightness is less than threshold value Mb2, weighting factors Zd1 to Zd4 are applied as weighting factors corresponding to lesion identification models A1, A2, C1, and C2, respectively.

[0173] Here, the weighting coefficients Xd1 to Xd4, Yd1 to Yd4, and Zd1 to Zd4 each take a value between 0 and 1. And, Xd1+Xd2+Xd3+Xd4=Yd1+Yd2+Yd3+Yd4=Zd1+Zd2+Zd3+Zd4=1.

[0174] Furthermore, the lower the average luminance, the higher the weighting coefficient of the lesion discrimination model A2 among the lesion discrimination models A1 and A2 is set. Specifically, Xd1 > Yd1 > Zd1, Xd2 < Yd2 < Zd2, and Xd1 > Xd2, Zd1 < Zd2. The magnitude relationship between Yd1 and Yd2 may be determined according to the set values of the threshold values Md1 and Md2.

[0175] On the other hand, for the lesion discrimination models C1 and C2, as described above, the weighting coefficient when the average luminance is an intermediate value may be set to a high value. Specifically, Yd3 > Xd3 > Zd3, Xd4 < Zd4 < Yd4 may be sufficient. The magnitude relationship between Yd3 and Yd4 may be determined according to the set values of the threshold values Md1 and Md2.

[0176] By using such a weighting coefficient, the detection result output unit 145 increases the weight of the probability value by the lesion detection model that is close to the conditions of the CT image group in which the conditions of the CT image used as the learning data are input among the lesion discrimination models A1, A2, C1, and C2, and can obtain the final lesion discrimination result. As a result, the lesion area can be accurately detected from a simple CT image.

[0177] Note that the processing functions of the apparatuses (for example, the information processing apparatus 1 and the diagnostic support apparatuses 100, 100a, and 100b) shown in the above embodiments can be realized by a computer. In that case, a program describing the processing contents of the functions that each apparatus should have is provided, and by executing that program on a computer, the above processing functions are realized on the computer. The program describing the processing contents can be recorded on a computer-readable recording medium. Examples of the computer-readable recording medium include a magnetic storage device, an optical disk, and a semiconductor memory. Examples of the magnetic storage device include a hard disk drive (HDD) and a magnetic tape. Examples of the optical disk include a CD (Compact Disc), a DVD (Digital Versatile Disc), and a Blu-ray Disc (BD, registered trademark).

[0178] 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.

[0179] 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]

[0180] 1. Information processing device 2 Storage section 3 Processing section 11, 12a, 12b, 13 Tomographic images 21,22 Lesion identification model

Claims

1. The computer classifying a plurality of first tomographic images obtained by photographing the inside of a plurality of first human bodies into a plurality of first tomographic image groups having different medical findings; a learning process for generating a plurality of first lesion identification models, each of which identifies whether each unit image region included in a tomographic image to be identified is a region of a specific lesion, by machine learning using different ones of the plurality of first tomographic image groups as learning data; calculating a first image feature amount based on a plurality of second tomographic images obtained by imaging the inside of a second human body; inputting the second tomographic images into each of the first lesion identification models, thereby obtaining, from each of the first lesion identification models, a probability that each of the unit image regions included in the second tomographic images is a region of the specific lesion; calculating an integrated value by integrating the probabilities acquired from the first lesion identification models for each of the unit image regions included in the second tomographic images based on the first image feature amount; a lesion detection process for detecting an area of ​​the specific lesion from each of the plurality of second tomographic images based on the integrated value; A lesion detection method is performed.

2. the first image feature amount is calculated based on pixel information of a region of a specific organ among image regions of the plurality of second tomographic images; The method of claim 1 .

3. The first image feature is an average luminance value in the region of the specific organ. The method of claim 2 .

4. the plurality of first tomographic images are classified based on a degree of fat accumulation in the specific organ; In calculating the integrated value, the integrated value is calculated by performing weighted addition on the probabilities obtained from each of the plurality of first lesion identification models, and the lower the average brightness, the higher the weighting coefficient is set for the probability from a first lesion identification model generated using the first tomographic image group in which a degree of fat accumulation is high, among the plurality of first lesion identification models. The method for detecting a lesion according to claim 3.

5. The learning process includes: calculating a second image feature of the same type as the first image feature for each second tomographic image group obtained by classifying the plurality of first tomographic images for each of the first human bodies; classifying the plurality of first tomographic images into a plurality of third tomographic image groups according to a range of the second image feature amount; generating a plurality of second lesion identification models for identifying whether each unit image region included in the tomographic image to be identified is a region of the specific lesion by machine learning using different ones of the plurality of third tomographic image groups as learning data; Including processing, The lesion detection process includes: inputting the second tomographic images into each of the second lesion identification models, thereby obtaining the probability for each of the unit image regions included in the second tomographic images from each of the second lesion identification models; Including processing, In calculating the integrated value, calculating an integrated value by integrating the probabilities acquired from each of the plurality of first lesion identification models and each of the plurality of second lesion identification models for each of the unit image regions included in the plurality of second tomographic images based on the first image feature amount; The method of claim 1 .

6. In the process of classifying the plurality of first tomographic images into the plurality of first tomographic image groups, the plurality of first tomographic images are classified into the first tomographic image group including a tomographic image of a normal finding and the first tomographic image group including a tomographic image of an abnormal finding; The learning process includes: generating a third lesion identification model that identifies whether or not each unit image region included in the tomographic image to be identified is a region of the specific lesion by machine learning using, as learning data, tomographic images of the normal findings, in which the second image feature value is included in a first range; a fourth lesion identification model that identifies whether each unit image region included in the tomographic image to be identified is a region of the specific lesion or not is generated by machine learning using, as learning data, tomographic images in which the second image feature value is included in a second range among the tomographic images of the abnormal findings; Processing The lesion detection process includes: inputting the plurality of second tomographic images into the third lesion identification model and the fourth lesion identification model, thereby obtaining the probability for each of the unit image regions included in the plurality of second tomographic images from each of the third lesion identification model and the fourth lesion identification model; Processing In calculating the integrated value, for each of the unit image regions included in the second tomographic images, the probabilities obtained from each of the first lesion identification models, each of the second lesion identification models, the third lesion identification model, and the fourth lesion identification model are integrated based on the first image feature to calculate the integrated value. The method for detecting a lesion according to claim 5.

7. On the computer, classifying a plurality of first tomographic images obtained by photographing the inside of a plurality of first human bodies into a plurality of first tomographic image groups having different medical findings; a learning process for generating a plurality of first lesion identification models, each of which identifies whether each unit image region included in a tomographic image to be identified is a region of a specific lesion, by machine learning using different ones of the plurality of first tomographic image groups as learning data; calculating a first image feature amount based on a plurality of second tomographic images obtained by imaging the inside of a second human body; inputting the second tomographic images into each of the first lesion identification models, thereby obtaining, from each of the first lesion identification models, a probability that each of the unit image regions included in the second tomographic images is a region of the specific lesion; calculating an integrated value by integrating the probabilities acquired from the first lesion identification models for each of the unit image regions included in the second tomographic images based on the first image feature amount; a lesion detection process for detecting an area of ​​the specific lesion from each of the plurality of second tomographic images based on the integrated value; A lesion detection program that executes the following:

Images