Image diagnosis device and image diagnosis method

Abstract

This image diagnostic device is provided with a processor that detects a first region, which is a lesion candidate region in a detection endoscopic image, using a learned model obtained by learning a region having a high degree of irregularity in a microvessel construction image and / or a surface microstructure in a learning endoscopic image, and determining a confidence level that the first region is a lesion region, and determining whether the first region is a lesion region based on the confidence level.

Description

Technical Field

[0001] This invention relates to an image diagnostic apparatus and an image diagnostic method. This application claims priority to U.S. Provisional Application No. 63 / 544,967, filed October 20, 2023, the contents of which are incorporated herein by reference. Background Technology

[0002] Previously, computer-aided image diagnosis (CAD) has been developed, which uses computers to analyze endoscopic images and assist surgeons in making diagnoses. For example, CAD has been designed to assist in the diagnosis of early gastric cancer. When early gastric cancer is detected by CAD, the surgeon is notified of its presence by presenting an emphasis display based on markers such as boxes on the endoscopic image.

[0003] In recent years, endoscopic devices equipped with narrow-band imaging (NBI, registered trademark) have been used in general clinical practice. In gastric magnification endoscopy using NBI, in addition to the microvascular pattern of the gastric mucosa, the microsurface pattern is also clearly visualized. Therefore, in addition to the microvascular pattern based on white light, the microsurface pattern can also be incorporated into the diagnostic criteria for early gastric cancer. Therefore, in gastric magnification endoscopy using NBI, a diagnostic algorithm that incorporates the microsurface pattern in addition to the microvascular pattern is proposed, namely the Magnifyingendoscopy simple diagnostic algorithm for early gastric cancer (MESDA-G) (Non-Patent Literature 1).

[0004] In addition, a system has been proposed that uses machine learning to extract features such as microvascular constructs and surface microstructures from NBI images to assist in the diagnosis of early gastric cancer (Patent Document 1).

[0005] Existing technical documents

[0006] Patent documents

[0007] Patent Document 1: International Publication No. 2020 / 224153

[0008] Non-patent literature

[0009] Non-patent literature 1: Muto M, Yao K, Kaise M et al. Magnifying endoscopy simplediagnostic algorithm for early gastric cancer (MESDA-G). Dig Endosc 2016; 28: 379-393. Summary of the Invention

[0010] The problem the invention aims to solve

[0011] However, in the diagnosis of early gastric cancer using MEDSA-G, higher accuracy in diagnosis is desired.

[0012] The present invention was made in view of the following circumstances, and its object is to provide an image diagnostic device and image diagnostic method that can achieve higher accuracy in the diagnosis of early gastric cancer using MEDSA~G.

[0013] Solution for solving the problem

[0014] To address the above problems, the present invention proposes the following solutions.

[0015] The image diagnostic apparatus according to the first aspect of the present invention includes a processor that uses a learned model, after learning regions with high irregularity in microvascular patterns and / or microsurface patterns in learning endoscope images, to detect lesion candidate regions, i.e., a first region, in detection endoscope images, and determines a confidence level that the first region is a lesion region, and determines whether the first region is a lesion region based on the confidence level.

[0016] The image diagnostic method of the second aspect of the present invention uses a learned model, after learning from the microvascular construct and / or the region with a high degree of irregularity in the surface microstructure in the learning endoscope image, to detect the lesion candidate region, i.e., the first region, in the detection endoscope image, and determines the confidence level of the first region as a lesion region, and determines whether the first region is a lesion region based on the confidence level.

[0017] The effects of the invention

[0018] The image diagnostic device and image diagnostic method of the present invention can achieve higher accuracy in the diagnosis of early gastric cancer using MEDSA-G. Attached Figure Description

[0019] Figure 1 This is a diagram illustrating the endoscope system according to the first embodiment.

[0020] Figure 2 This is a functional block diagram of the endoscope system.

[0021] Figure 3 This is a diagram showing the input and output of the learned model of the endoscope system.

[0022] Figure 4 It refers to the endoscopic images used in the learning of the model after it has been learned.

[0023] Figure 5 This is a diagram showing the candidate lesion region, i.e., the first region, within the camera image (detection endoscope image) detected using the learned model.

[0024] Figure 6 This is a diagram showing an example of a composite image.

[0025] Figure 7 This is a flowchart illustrating the operation of the endoscope system.

[0026] Figure 8 This is a diagram showing the input and output of the learned model of the endoscope system according to the second embodiment.

[0027] Figure 9 This is a diagram showing the input and output of the learned model of the endoscope system according to the third embodiment. Detailed Implementation

[0028] (First Implementation)

[0029] Reference Figures 1 to 7 To illustrate an embodiment of the present invention, an endoscope system 500 is described.

[0030] [Endoscopic System 500]

[0031] Figure 1 This is a diagram showing the endoscope system 500.

[0032] Endoscopic system 500 is a system capable of performing gastric magnification endoscopy using NBI (Narrow Band Imaging, registered trademark). Endoscopic system 500 includes an endoscope 100, an image processor 200, a light source 300, and a display device 400. The image processor 200 and the light source 300 can be integrated into one device (image control device).

[0033] The light source device 300 has a light source 310 such as an LED, and controls the light source to control the amount of illumination light transmitted to the endoscope 100 via the light guide 161.

[0034] Display device 400 is a device that displays images generated by image processor device 200, various information related to endoscope system 500, etc. Display device 400 is, for example, a liquid crystal monitor or a head-mounted display.

[0035] [Endoscope 100]

[0036] Endoscope 100 is, for example, a device for observing and treating the body of a patient lying on an examination table T. Endoscope 100 includes an elongated insertion part 110 that is inserted into the patient's body, an operating part 180 connected to the base of the insertion part 110, and a universal flexible cable 190 extending from the operating part 180.

[0037] The insertion part 110 has a front end portion 120, a flexible bending portion 130, and a long, flexible tube portion 140. The front end portion 120, the bending portion 130, and the flexible tube portion 140 are connected sequentially from the front end side. The flexible tube portion 140 is connected to the operation part 180.

[0038] Figure 2 This is a functional block diagram of the endoscopic system 500.

[0039] The front end 120 has a camera unit 150 and a lighting unit 160.

[0040] The camera unit 150 includes an optical system, an imaging element that converts light signals into electrical signals, and an analog-to-digital converter circuit that converts the analog signals output by the imaging element into digital signals. The camera unit 150 is equipped with a zoom function, enabling switching between normal observation and magnified observation. The camera unit 150 captures an image of a subject and generates an image signal. The image signal is acquired by the image processor device 200 via the image signal cable 151.

[0041] The illumination unit 160 illuminates the subject with illumination light (white light) transmitted by the light guide 161 or with special light used in NBI. The light guide 161 passes through the insertion part 110, the operation part 180, and the universal flexible cable 190 to connect to the light source device 300. In addition, the illumination unit 160 may also have light sources such as LEDs, optical elements such as phosphors with wavelength conversion functions, etc.

[0042] The operation unit 180 handles operations related to the endoscope 100. The operation unit 180 includes a bending knob 181 for controlling the bending section 130, an air / water supply button 182, a suction button 183, a release button 184, and a zoom lever (not shown). The bending knob 181 is a rotary handle that bends the bending section 130. Operations input to the air / water supply button 182, the suction button 183, and the release button 184 are received by the image processor device 200. The release button 184 is used to input an operation to save the image captured by the imaging unit 150. The zoom lever is a lever for switching between normal observation and magnified observation.

[0043] A general-purpose flexible cable 190 connects the endoscope 100 to the image processor device 200. The general-purpose flexible cable 190 is a cable through which camera signal cables 151, light guides 161, etc., pass.

[0044] [Image Processor Device 200]

[0045] like Figure 2 As shown, the image processor device (endoscopic diagnostic device) 200 includes an image acquisition unit 220, an image recording unit 230, a detection unit 240, and an image synthesis unit 290.

[0046] The image processor device 200 is a computer equipped with a processor such as a CPU, memory, and a recording unit, capable of executing programs. The functions of the image processor device 200 are implemented by the processor executing programs. At least some of the functions of the image processor device 200 can also be implemented using dedicated logic circuits mounted on an ASIC or FPGA.

[0047] The image processor device 200 may also have structures other than a processor, memory, and recording unit. For example, the image processor device 200 may also have part or all of an image processing unit that performs image processing and image recognition processing. By also having an image processing unit, the image processor device 200 can perform specific image processing and image recognition processing at high speed. The image processing unit may also be a processor located on a cloud server connected via the Internet.

[0048] The recording unit is a non-volatile recording medium that stores the aforementioned program and the data required to execute the program. The recording unit may be composed of, for example, writable non-volatile memory such as floppy disks, optical disks, ROM, and flash memory; portable media such as CD-ROMs; and storage devices such as hard disks and SSDs built into a computer system. The recording unit may also be a storage device located on a cloud server connected via the Internet.

[0049] The aforementioned program can be provided, for example, by a computer-readable recording medium such as flash memory. The program can also be transmitted from the computer holding the program to a memory or recording unit via a transmission medium or by transmission waves within the transmission medium. The "transmission medium" for transmitting the program is a medium capable of transmitting information. Media capable of transmitting information include networks (communication networks) such as the Internet and communication lines (communication lines) such as telephone lines. The aforementioned program can also implement a portion of the aforementioned functions. Furthermore, the aforementioned program can also be a differential file (differential program). The aforementioned functions can also be implemented through a combination of a program already recorded in the computer and a differential program.

[0050] At least a portion of the image processor device 200 may also be a device separate from the image processor device 200. The separate device may also be a computing device located on a cloud server connected via the Internet.

[0051] The image acquisition unit 220 acquires a camera signal from the camera unit 150 of the endoscope 100 via the camera signal cable 151. The image acquisition unit 220 performs camera signal processing on the camera signal acquired from the camera unit 150 to sequentially acquire camera images D. The image acquisition unit 220 outputs the acquired camera image D to the image synthesis unit 290. Additionally, the acquired camera image D is output to the detection unit 240 via the image recording unit 230.

[0052] The image recording unit 230 is part of the aforementioned recording unit and is a non-volatile recording medium. The image recording unit 230 may also be part of the aforementioned memory and is a volatile recording medium. The image recording unit 230 records multiple transmitted camera images D.

[0053] Multiple camera images D (time-series images) input in chronological order are recorded in the image recording unit 230. When the recording capacity of the image recording unit 230 is insufficient, the earliest camera image D is deleted. The multiple camera images D recorded in the image recording unit 230 can be consecutive camera images D, or they can be camera images D obtained by removing multiple frames at intervals from consecutive frames.

[0054] The detection unit 240 detects candidate lesion regions, i.e., the first region R1, within the camera image (the endoscope image for detection) D based on the learned model 250. The learned model 250 is, for example, a semantic segmentation model. However, the learned model 250 is not limited to this and can also be other machine learning models capable of outputting information from the input image.

[0055] The detection unit 240 can also detect the first region R1 only from the camera image (detection endoscope image) D saved based on the operation input to the release button 184.

[0056] Figure 3 This is a diagram showing the input and output of the learned model 250.

[0057] After learning, the model (inference model) 250 was trained, and the region with a high degree of irregularity in the microvascular construct and / or surface microstructure in the Magnifying endoscopy simple diagnostic algorithm for early gastric cancer (MESDA-G) was designated as the first region R1.

[0058] Specifically, after learning, Model 250 learned that regions in the camera image D with one or more of the following characteristics—uniform shape, asymmetrical distribution, and irregular arrangement—were regions with a high degree of irregularity in the microvascular construct image.

[0059] Specifically, after learning, model 250 learned regions in the camera image D that have one or more of the following characteristics: uneven shape, asymmetrical distribution, or irregular arrangement, as regions with a high degree of irregularity in the surface microstructure.

[0060] Figure 4 This refers to an image labeled with label L, which is a region in the learning endoscope image included in the teaching data used in the learning of the completed model 250 that the doctor diagnosed as having a high degree of irregularity. Label L can also be separated into another image, also known as a mask image. By using a variety of learning endoscope images in the learning process, a learned model 250 with high diagnostic accuracy can be generated.

[0061] The degree of irregularity can be binary or multi-valued. In the case of multi-valued irregularity, different labels L are assigned to the endoscopic images based on the degree of irregularity. In this case, after training, model 250 can detect regions with high irregularity as the first region R1 and also detect the degree of irregularity.

[0062] The detection unit 240 determines the confidence level that the first region R1 is a lesion region. The confidence level is the degree of confidence in estimating that the first region R1 is cancer. In this embodiment, the higher the degree of irregularity in the microvascular construct and / or surface microstructure, the higher the confidence level for determining it as the first region R1.

[0063] Confidence levels can be graded into multiple levels. These graded confidence levels include, for example, a high confidence level (first confidence level) where areas with high irregularity have an extremely high probability of being cancer, a moderate confidence level (second confidence level) where areas with high irregularity have a high probability of being cancer, and a low confidence level (third confidence level) where areas with high irregularity have a relatively low probability of being cancer.

[0064] The detection unit 240 can output a determination result J on whether the first region R1 is cancerous based on the confidence level.

[0065] Figure 5 This is a diagram showing the candidate lesion region, i.e., the first region R1, within the camera image (detection endoscope image) D detected using the learned model 250. The learned model 250 is able to detect the first region R1 and determine the confidence level of the first region R1. The confidence level output from the learned model 250 can be used as an indicator for doctors and others who have reviewed the diagnostic information E (first region R1, confidence level, determination result) output by the detection unit 240 to determine that the first region R1 is cancer.

[0066] The detection unit 240 can also detect the demarcation line (DL) between lesions and non-lesions in the Magnifying endoscopy simple diagnostic algorithm for early gastric cancer (MESDA-G), which is used as a diagnostic algorithm for early gastric cancer, based on the second learned model. The demarcation line DL is a clear boundary line that appears after the regular microvascular structure or surface microstructure around the lesion has disappeared when observed with a magnified endoscope. The detection unit 240 detects the region with a clear demarcation line DL as a second region R2.

[0067] The detection unit 240 can also detect regions with irregularity levels above a first threshold as first regions R1, and regions with irregularity levels below the first threshold and above a second threshold that is lower than the first threshold as second regions R2.

[0068] The detection unit 240 can detect regions with a high confidence level (first confidence level) as first regions R1, and regions with a medium confidence level (second confidence level) as second regions R2.

[0069] The detection unit 240 outputs the detected diagnostic information E (first region R1, second region R2, confidence level, judgment result J) to the image synthesis unit 290.

[0070] Figure 6This is a diagram illustrating an example of a synthesized image S.

[0071] The image compositing unit (display unit) 290 generates a composite image S containing the captured image D and diagnostic information E related to the first region R1 detected by the detection unit 240. The image compositing unit 290 sequentially creates composite images S corresponding to the captured images D sequentially generated by the image acquisition unit 220 and outputs them to the display device 400. The display device 400 sequentially displays the received composite images S.

[0072] exist Figure 6 In the illustrated composite image S, markers such as different colors and patterns are used to emphasize a first region R1, which represents a high confidence level (first confidence level), and a second region R2, which represents a medium confidence level (second confidence level). The image compositing unit 290 changes the display method between the first region R1 and the region outside it.

[0073] The image synthesis unit 290 emphasizes the display of the first region R1 when the area ratio of the first region R1 to the area of ​​the second region R2 is less than a threshold, and does not emphasize the display of the first region R1 when the area ratio is greater than or equal to the threshold. This prevents the small first region R1 from being overlooked.

[0074] In addition, Figure 6 The illustrated composite image S displays a historical image H of the captured image highlighted above. The displayed historical image H is, for example, a captured image D saved based on the operation input to the release button 184.

[0075] [Operation of Endoscopic System 500]

[0076] Next, the operation (auxiliary information generation method) of the endoscope system 500 will be explained. Specifically, a surgical procedure using the endoscope system 500 to observe the gastric wall using a gastric magnification endoscope with NBI will be described. The following will follow... Figure 7 The flowchart shown illustrates the operation of the endoscope system 500.

[0077] <Step S110>

[0078] In step S110, the image acquisition unit 220 acquires a photographic image (endoscopic image for detection) D of the gastric mucosa with a clear image of the microvascular structure and surface microstructure by observation using illumination light (white light) and NBI. The endoscope system 500 then executes step S120.

[0079] <Step S120>

[0080] In step S120, the detection unit 240 detects the first region R1 in the camera image D. If the first region R1 is detected, the endoscope system 500 then executes step S130.

[0081] <Step S130>

[0082] In step S130, the detection unit 240 determines the confidence level of the detected first region R1. In this embodiment, the higher the degree of irregularity in the microvascular construct and / or surface microstructure, the higher the confidence level for determining it as the first region R1. The endoscope system 500 then executes step S140.

[0083] <Step S140>

[0084] In step S140, the detection unit 240 detects the second region R2 as needed. The endoscope system 500 then executes step S150.

[0085] <Step S150>

[0086] In step S150, as Figure 6 As shown, the image synthesis unit 290 generates a composite image S containing the captured image D and the diagnostic information E (first region R1, second region R2, confidence level, and judgment result J) generated by the detection unit 240, and outputs it to the display device 400. The display device 400 displays the received composite image S. The endoscope system 500 then executes step S160.

[0087] <Step S160>

[0088] In step S160, the detection unit 240 determines whether the surgery has ended. If the detection unit 240 determines that the surgery has not ended, it executes the steps after step S110. If the detection unit 240 determines that the surgery has ended, it executes step S170 to terminate the procedure. Figure 7 The control flow is shown.

[0089] The endoscope system 500 according to this embodiment can achieve higher accuracy in the diagnosis of early gastric cancer using MEDSA-G. When a diagnostic algorithm for region discrimination, such as semantic segmentation, is applied to the photographic image (endoscopic image for detection) D, even if a very small lesion area is detected, it is easily judged as a false detection (noise). However, the endoscope system 500 detects a first region R1 as a candidate lesion region and determines the confidence level for the first region R1. Therefore, even if the detected first region R1 is a very small region, if the confidence level is high, the endoscope system 500 will not judge it as a false detection (noise).

[0090] The first embodiment of the present invention has been described in detail above with reference to the accompanying drawings. However, the specific structure is not limited to this embodiment, and design changes are also included without departing from the spirit of the present invention. Furthermore, the constituent elements shown in the embodiments and variations can be appropriately combined to form a configuration.

[0091] (Second Implementation)

[0092] Reference Figure 8 The second embodiment of the present invention will now be described. In the following description, structures that are common to those already described will be labeled with the same reference numerals and repeated descriptions will be omitted.

[0093] The endoscope system 500B involved in the second embodiment, like the endoscope system 500 of the first embodiment, is a system capable of performing gastric magnification endoscopy observation using NBI. For example... Figure 2 As shown, the endoscope system 500B includes an endoscope 100, an image processor device 200B, a light source device 300, and a display device 400.

[0094] The image processor device 200B includes an image acquisition unit 220, an image recording unit 230, a detection unit 240B, and an image synthesis unit 290.

[0095] Figure 8 This is a diagram showing the input and output of the learned Model 250B.

[0096] The detection unit 240B detects the first region R1, a candidate lesion region, within the camera image (endoscopic image for detection) D based on the learned model 250B. Compared to the learned model 250 of the first embodiment, the learned model (inference model) 250B can also output a confidence level.

[0097] In the training data used for Model 250B, regions in the endoscopic images included in the training data were labeled L for areas diagnosed by the physician as having a high degree of irregularity, and a confidence level was assigned to the label L. That is, the training endoscopic images were assigned to the physician's diagnosis of irregularities in microvascular constructs and / or surface microstructures, and the physician's confidence level for that diagnosis. Furthermore, the confidence level assigned by the physician was not only based on information obtained from the images used as the subjects of the assignment, but sometimes also considered other images from the endoscopic examination, the patient's medical history, age, consultation results, and other information to make a comprehensive judgment and assignment.

[0098] Specifically, after learning Model 250B, it outputs, for example, a score and reliability estimate of similarity to regions with high irregularity, on a pixel-by-pixel basis. Regions with a score above a certain level and a reliability estimate are then used; the higher the reliability estimate, the higher the confidence level of that region.

[0099] According to the endoscope system 500B of this embodiment, higher accuracy in the diagnosis of early gastric cancer using MEDSA-G can be achieved. By using a learned model 250B that has been trained to correspond to a doctor's confidence level in diagnosing microvascular structures and / or irregularities in surface microstructures, the learned model 250B can output a confidence level that simulates the doctor's judgment. Therefore, doctors and others who review the diagnostic information E (first region R1, confidence level, judgment result) output by the detection unit 240B can perform early gastric cancer diagnosis with higher accuracy.

[0100] The second embodiment of the present invention has been described in detail above with reference to the accompanying drawings. However, the specific structure is not limited to this embodiment, and design changes are also included without departing from the spirit of the present invention. Furthermore, the constituent elements shown in the embodiments and variations can be appropriately combined to form a configuration.

[0101] (Third Implementation)

[0102] Reference Figure 9 The second embodiment of the present invention will now be described. In the following description, structures that are common to those already described will be labeled with the same reference numerals and repeated descriptions will be omitted.

[0103] The endoscope system 500C involved in the third embodiment, like the endoscope system 500 of the first embodiment, is a system capable of performing gastric magnification endoscopy observation using NBI. For example... Figure 2 As shown, the endoscope system 500C includes an endoscope 100, an image processor device 200C, a light source device 300, and a display device 400.

[0104] The image processor device 200C includes an image acquisition unit 220, an image recording unit 230, a detection unit 240C, and an image synthesis unit 290.

[0105] Figure 9 This is a graph showing the input and output of the learned Model 250C.

[0106] The detection unit 240C detects the first region R1, a candidate lesion region, within the camera image (detection endoscope image) D, based on the learned model 250C. Compared to the learned model 250 of the first embodiment, the learned model (inference model) 250C uses confidence levels instead of the degree of irregularity.

[0107] In the training data for Model 250C, regions in the endoscopic images used for learning were labeled L for areas diagnosed by the physician as having a high degree of irregularity, and each label L was assigned a confidence level by the physician. The confidence level was multi-valued, and the confidence level was assigned to each region based on the physician's level of confidence in the diagnosis.

[0108] Specifically, after learning, model 250B estimates the confidence level, for example, on a pixel-by-pixel basis. Detection unit 240C detects regions estimated to have high confidence levels as first regions R1. Detection unit 240C can detect regions with the same confidence level as a single first region R1, or it can detect multiple regions with different confidence levels as a single first region R1. In the latter case, detection unit 240C can set the highest confidence level as the confidence level of the first region R1.

[0109] According to the endoscope system 500C of this embodiment, higher accuracy can be achieved in the diagnosis of early gastric cancer using MEDSA-G. By using a learned model 250C that has learned the confidence level of a doctor corresponding to the doctor's diagnosis of microvascular structures and / or irregularities in surface microstructures, the learned model 250C can output a confidence level that simulates the doctor's judgment. Therefore, doctors and others who review the diagnostic information E (first region R1, confidence level, judgment result) output by the detection unit 240C can perform early gastric cancer diagnosis with higher accuracy.

[0110] The third embodiment of the present invention has been described in detail above with reference to the accompanying drawings. However, the specific structure is not limited to this embodiment, and design changes are also included without departing from the spirit of the present invention. In addition, the constituent elements shown in the embodiments and variations can be appropriately combined to form a configuration.

[0111] [Industry availability]

[0112] This invention can be applied to endoscope systems, etc.

[0113] Explanation of reference numerals in the attached figures

[0114] 500: Endoscopic system; 400: Display device; 100: Endoscope; 200: Image processor device (endoscopic diagnostic device); 220: Image acquisition unit; 230: Image recording unit; 240: Detection unit; 250: Learned model (inference model); 290: Image synthesis unit (display unit); 300: Light source device; D: Camera image (endoscopic image for detection); E: Diagnostic information; H: Historical record image; J: Judgment result; L: Label; R1: First region; R2: Second region; S: Synthesized image.

Claims

1. An image diagnostic device, comprising a processor, The processor performs the following processing: The learned model, which is trained on regions with high irregularity in the microvascular constructs and / or surface microstructures in the learning endoscope images, is used to detect the lesion candidate regions, i.e., the first region, in the detection endoscope images. Determine the confidence level at which the first region is a lesion region; as well as The determination of whether the first region is a lesion region is based on the confidence level.

2. The image diagnostic device according to claim 1, wherein, The processor performs the following processing: The region with a high degree of irregularity within the endoscopic image used for detection is identified as the first region; and The higher the degree of irregularity of the microvascular construct and / or the surface microstructure in the first region, the higher the confidence level of the first region.

3. The image diagnostic device according to claim 2, wherein, The learned model is a model that learns regions in the endoscopic image of the detection device that have one or more of the following characteristics: uneven shape, asymmetrical distribution, and irregular arrangement, as regions with a high degree of irregularity in the microvascular construct image.

4. The image diagnostic device according to claim 2, wherein, The learned model is a model that learns from regions in the endoscopic image of the detection device that have one or more of the following characteristics: uneven shape, asymmetrical distribution, and irregular arrangement, as regions with a high degree of irregularity in the surface microstructure.

5. The image diagnostic device according to claim 1, wherein, The degree of irregularity is multivalued.

6. The image diagnostic apparatus according to claim 1, wherein, The learned model is a model that, in addition to learning the regions with high irregularity in the endoscopic images used for learning, also learns the confidence levels of the regions with high irregularity. The confidence level included in the teaching data used in the learning of the completed model is based on the doctor's confidence level in diagnosing the areas with a high degree of irregularity. The processor uses the learned model to detect the first region and determines the confidence level for the first region.

7. The image diagnostic apparatus according to claim 6, wherein, The doctor's judgment is a comprehensive assessment that takes into account at least one of the following information, in addition to the information obtained from the endoscopic images used for testing: other endoscopic images from the endoscopic examination, the patient's medical history, age, and consultation results.

8. The image diagnostic apparatus according to claim 1, wherein, The confidence levels are divided into multiple levels. The confidence levels determined by the classification include: The region with a high degree of irregularity represents a first confidence level indicating a very high probability of the lesion region; and The area with a high degree of irregularity represents a second confidence level indicating a high probability of the lesion area.

9. The image diagnostic apparatus according to claim 8, wherein, The processor detects a second region that includes the first region and has a clear boundary line. When the confidence level of the first region is above a threshold, the processor determines the second region as the lesion region.

10. The image diagnostic apparatus according to claim 9, wherein, The processor detects regions with irregularities exceeding a first threshold as the first region. The processor detects regions with irregularities that are less than the first threshold and greater than a second threshold that is lower than the first threshold as the second region.

11. The image diagnostic apparatus according to claim 8, wherein, The region that is detected as the first confidence level is detected as the first region. The region that is detected as the second confidence level is detected as a second region with a clear boundary line.

12. The image diagnostic apparatus according to claim 9, wherein, The image diagnostic device includes a display unit that emphasizes the endoscopic image for testing and the first region.

13. The image diagnostic apparatus according to claim 12, wherein, The display unit changes its display method between the first region and the region outside it.

14. The image diagnostic apparatus according to claim 13, wherein, When the area ratio of the first region to the area of ​​the second region is less than a threshold, the display unit emphasizes the first region. If the area ratio is above the threshold, the first region is not highlighted.

15. An image diagnostic method, wherein, The learned model, which is trained on regions with high irregularity in the microvascular constructs and / or surface microstructures in the learning endoscope images, is used to detect the lesion candidate regions, i.e., the first region, in the detection endoscope images. Determine the confidence level at which the first region is a lesion region; as well as The determination of whether the first region is a lesion region is based on the confidence level.

16. The image diagnostic method according to claim 15, wherein, The region with a high degree of irregularity within the endoscopic image used for detection is identified as the first region. The higher the degree of irregularity of the microvascular construct and / or the surface microstructure in the first region, the higher the confidence level of the first region.

17. The image diagnostic method according to claim 15, wherein, The learned model is a model that, in addition to learning the regions with high irregularity in the endoscopic images used for learning, also learns the confidence levels of the regions with high irregularity. The confidence level included in the teaching data used in the learning of the completed model is based on the doctor's confidence level in diagnosing the areas with a high degree of irregularity. In the image diagnostic method, the learned model is used to detect the first region, and the confidence level for the first region is determined.

18. An image diagnostic method, wherein, The learned model, trained on confidence levels of microvascular constructs and / or surface microstructures within endoscopic images used for training, is used to detect lesion candidate regions, i.e., the first region, within endoscopic images used for detection. Determine the confidence level of the first region. The determination of whether the first region is a lesion region is based on the confidence level.

Images