Endoscopic image diagnosis assistance processor, operation method for endoscopic image diagnosis assistance processor, and program for endoscopic image diagnosis assistance processor

Abstract

Provided is an endoscopic image diagnosis assistance processor (2) that stably and efficiently identifies an organ under observation. The endoscopic image diagnosis assistance processor (2) comprises: a site inference unit (13) that uses a first model to infer a site confidence level that an endoscopic image is an image of a site; a site identification unit (14) that uses the site confidence level to identify a site under observation in the endoscopic image; an organ identification unit (15) that identifies a first organ candidate corresponding to the site under observation; an organ inference unit (17) that uses a second model to infer a second organ candidate corresponding to the endoscopic image; and an organ determination unit (18) that determines the first organ candidate as the organ under observation in the endoscopic image when the first organ candidate and the second organ candidate are the same, and does not determine the organ under observation or determines the organ under observation on the basis of a predetermined condition when the first organ candidate and the second organ candidate are different.

Description

Endoscopic image diagnosis support processor, method of operating an endoscopic image diagnosis support processor, and program for an endoscopic image diagnosis support processor

[0001] Embodiments of the present invention relate to an endoscopic image diagnosis support processor capable of efficiently identifying an organ being observed, a method of operating an endoscopic image diagnosis support processor capable of efficiently identifying an organ being observed, and a program for an endoscopic image diagnosis support processor capable of efficiently identifying an organ being observed.

[0002] Computer-aided detection (CADe: Computer-Aided Detection) indicating the position of a lesion candidate and computer-aided diagnosis (CADx: Computer-Aided Diagnosis) indicating discrimination information of a lesion candidate are known for an image captured using an endoscope (hereinafter, CADe and CADx are collectively referred to as "CAD"). In CAD, for example, a machine learning model that has performed deep learning or the like using a plurality of images as teacher data is used.

[0003] International Publication No. WO2022 / 181748 discloses a technique for generating a plurality of models each adapted to a plurality of organs by learning using images of each of the plurality of organs. For example, a plurality of types of models such as a pharyngeal model learned using only pharyngeal images and an esophageal model learned using only esophageal images are created. Then, by using a model adapted to the organ being observed with an endoscope from among the plurality of types of models, the accuracy as CAD can be improved.

[0004] Japanese Patent Application Laid-Open No. 2007-151809 discloses a receiving device that detects that the organ being photographed by a capsule endoscope has changed and receives data from the capsule endoscope.

[0005] However, it is not easy to identify the organ being observed with an endoscope when the imaging conditions are not good, and in some cases, there is a risk that an incorrect organ may be identified. Also, even when the imaging conditions are good, there are cases where the organ being observed cannot be correctly identified 100%.

[0006] International Publication WO2022 / 181748, Japanese Patent Publication No. 2007-151809

[0007] Embodiments of the present invention aim to provide an endoscopic image diagnostic support processor for efficiently identifying an organ being observed with an endoscope, a method for operating the endoscopic image diagnostic support processor for efficiently identifying an organ being observed with an endoscope, and a program for the endoscopic image diagnostic support processor that enables a computer to efficiently identify an organ being observed with an endoscope.

[0008] The endoscopic image diagnostic support processor according to an embodiment of the present invention comprises: a site inference unit that, for a plurality of endoscopic images input in succession, uses a first model trained with a plurality of images of a plurality of sites as training data to infer a site confidence level for each endoscopic image that is an image of the plurality of sites; a site identification unit that identifies an observation site of the endoscopic image based on predetermined conditions using the site confidence level; an organ identification unit that identifies a first organ candidate corresponding to the observation site; an organ inference unit that infers at least one second organ candidate corresponding to each endoscopic image using at least one second model trained with a plurality of images of at least one organ as training data; and an organ identification unit that, if the first organ candidate and the second organ candidate are the same, determines the first organ candidate as the observation organ of the endoscopic image, and if the first organ candidate and the second organ candidate are different, either does not determine the observation organ or determines the observation organ based on predetermined conditions.

[0009] The operation method of the endoscopic image diagnostic support processor according to an embodiment of the present invention involves, for a plurality of endoscopic images that are input in succession, using a first model that has been trained using multiple images of multiple body parts as training data, inferring a degree of confidence that each endoscopic image is an image of one of the multiple body parts, identifying the observation site of the endoscopic image and identifying a first organ candidate corresponding to the observation site based on predetermined conditions using the degree of confidence, inferring at least one second organ candidate corresponding to each endoscopic image using at least one second model that has been trained using multiple images of at least one organ as training data, and if the first organ candidate and the second organ candidate are the same, determining the first organ candidate as the observation organ of the endoscopic image, and if the first organ candidate and the second organ candidate are different, either not determining the observation organ or determining the observation organ based on predetermined conditions.

[0010] The program for the endoscopic image diagnostic support processor according to an embodiment of the present invention, for a plurality of endoscopic images that are input sequentially, uses a first model that has been trained using multiple images of multiple body parts as training data to infer the confidence level for each endoscopic image to be an image of one of the multiple body parts, identifies the observation site of the endoscopic image and identifies a first organ candidate corresponding to the observation site based on predetermined conditions using the confidence level, uses at least one second model that has been trained using multiple images of at least one organ as training data to infer at least one second organ candidate corresponding to each endoscopic image, if the first organ candidate and the second organ candidate are the same, confirms the first organ candidate as the observation organ of the endoscopic image, and if the first organ candidate and the second organ candidate are different, either does not confirm the observation organ or causes the computer to execute a process to confirm the observation organ based on predetermined conditions.

[0011] According to embodiments of the present invention, it is possible to provide an endoscopic image diagnostic support processor that efficiently identifies an organ being observed with an endoscope, a method for operating the endoscopic image diagnostic support processor that efficiently identifies an organ being observed with an endoscope, and a program for the endoscopic image diagnostic support processor that allows a computer to efficiently identify an organ being observed with an endoscope.

[0012] Figure 1 is a configuration diagram of the endoscopic image diagnostic support processor of the first embodiment. Figure 2 is a flowchart of the operation method of the endoscopic image diagnostic support processor of the first embodiment. Figure 3 is an example of the screen of the display device of the embodiment. Figure 4 is a configuration diagram of the endoscopic image diagnostic support processor of the second embodiment. Figure 5 is a flowchart of the operation method of the endoscopic image diagnostic support processor of the second embodiment.

[0013] Embodiments of the present invention will be described below with reference to the drawings. The drawings based on the embodiments are schematic. The illustration and reference numerals of some components have been omitted.

[0014] <First Embodiment> As shown in Figure 1, the endoscopic image diagnostic support processor 2 (hereinafter referred to as "processor 2") of this embodiment, together with the endoscope 9, display device 30, notification unit 31, and storage device 32, constitutes the endoscopic system 1.

[0015] The camera unit 8 of the endoscope 9 captures images of the inside of the subject's body. The endoscope 9 outputs image data at a predetermined frame rate. The user can capture still images, for example, by operating the operation buttons 7 on the control panel (not shown) of the endoscope 9. The endoscope 9 may be either a flexible or rigid endoscope.

[0016] The processor 2 includes an observed organ determination device 10 (hereinafter referred to as "determination device 10") and a lesion detection unit 20.

[0017] The determination device 10 includes a CPU 11, an image input unit 12, a site inference unit 13, a site identification unit 14, an organ identification unit 15, a memory 16, an organ inference unit 17, and an organ determination unit 18. The processor 2 is also connected to the server 40.

[0018] The CPU 11 is a system controller that controls the entire processor 2. The image input unit 12 processes images taken by the endoscope 9 at a first frame rate. Alternatively, the image data from the endoscope 9 may be processed by an image processing processor (not shown) or a server 40 before being input to the image input unit 12.

[0019] The site inference unit 13 infers the confidence level for each of the multiple sites in the endoscopic image, whether it is an inappropriate image, or whether it is an undeterminable image. For example, the site inference unit 13 is a first AI processing unit that uses a first model, which has been trained using multiple images of multiple sites as training data, to infer the site confidence level RP for each of the multiple sites in the input endoscopic image using the first model.

[0020] The site identification unit 14 identifies an observation site, which is a part of the endoscopic image, from among multiple sites based on predetermined conditions using the site confidence score RP. For example, the site identification unit 14 identifies the site with the highest site confidence score RP as the observation site.

[0021] The organ identification unit 15 identifies a first candidate organ to be observed that corresponds to the observation site. The memory 16 stores various data, such as the correspondence between the observation site and the observed organ.

[0022] The organ inference unit 17 is a second AI processing unit that uses a second model, which has been trained using multiple images of multiple organs as training data, to infer a second organ candidate corresponding to each endoscopic image.

[0023] If the first organ candidate and the second organ candidate are the same, the organ determination unit 18 determines the first organ candidate as the organ to be observed in the endoscopic image. If the first organ candidate and the second organ candidate are different, the organ determination unit 18 either does not determine the observed organ or determines the observed organ based on predetermined conditions.

[0024] For example, the organ inference unit infers the organ with the highest organ confidence score OP as the second organ candidate.

[0025] The lesion detection unit 20 is a second AI processing unit that infers lesions in images using multiple third models (organ CADs) that have been trained using multiple lesion images for multiple organs as training data.

[0026] The lesion detection unit 20 includes a nasal cavity CAD (third model 1) 21, an oral cavity CAD (third model 2) 22, a pharyngeal CAD (third model 3) 23, an esophageal CAD (third model 4) 24, a gastric CAD (third model 5) 25, a duodenal CAD (third model 6) 26, and an ON / OFF control unit 29 that controls the activation / deactivation of each model. Each model may also have its own ON / OFF control unit.

[0027] Furthermore, at least one of the multiple configurations of the processor 2 may consist of internal circuits of semiconductor elements processed by software, or dedicated hardware circuits, or it may include both internal circuits of semiconductor elements and dedicated hardware circuits. At least one of the functional parts of the processor 2 may also be a component of the server 40. For example, the lesion detection unit 20 may be an operating unit of the server 40.

[0028] For example, the body part inference unit 13 may include an FPGA (Field Programmable Gate Array) with hardware circuits and a GPU (Graphics Processing Unit) with software circuits. Also, the body part identification unit 14, organ identification unit 15, organ determination unit 18, etc., may be part of a CPU 11 that processes data using software.

[0029] The display device 30, which displays the endoscopic image, is a monitor such as a liquid crystal display. The notification unit 31 informs the user of information from the processor 2 using sound, light, etc. The notification unit 31 may be part of the display device 30.

[0030] The storage device 32 is a non-temporary storage device (e.g., a magnetic disk, optical disk, or HDD) that stores the program of the processor 2. The storage device 32 may also be configured as a server 40 connected to the processor 2 via an internet connection or the like.

[0031] The organ determination unit 18 determines the observed organ from the first organ candidate and the second organ candidate identified by different inference algorithms, based on predetermined conditions.

[0032] As will be described later, the determination device 10 of the processor 2 can efficiently and stably identify the organ being observed in the endoscopic image. The third CAD model of the lesion detection unit 20 is created for each of multiple organs. Therefore, the processor 2 has good model learning efficiency and high inference reliability.

[0033] <Processor Operation Method> The operation method of processor 2 will be explained according to the flowchart in Figure 2.

[0034] <Step S10> Image Input The endoscopic system 1 of this embodiment performs an examination of the upper gastrointestinal tract. The insertion part of the endoscope 9 is inserted through the oral cavity or nasal cavity of the subject and enters the pharynx. For example, multiple endoscopic images from the endoscope 9 or the image processing processor are input continuously to the image input unit 12 of the processor 2 at a first frame rate (e.g., 120 fps: 120 frames / second). Processing is not performed if no images are input.

[0035] The image input unit 12 performs image processing on images directly input from the endoscope 9 (for example, RAW images) for display on the display device 30. If an image processed by an image processing processor is input, no image processing is required.

[0036] <Step S20> Site Inference The site inference unit 13 infers the site confidence RPN (N=1 to n) that the input endoscopic image is an image of one of n sites. The value of confidence P (RP) is in the range of (0 to 1). (Confidence P=0) means a probability of 0%, and (Confidence P=1) means a probability of 100%.

[0037] The machine learning model of the AI processing unit is a deep neural network (DNN: Deep Neural Network) that has multiple hidden layers and performs deep learning. The model may also be a convolutional neural network (CNN: Convolution Neural Network), an R-CNN (Regions with CNN features) using a CNN, or a fully convolutional network (FCN), etc.

[0038] The first model is trained using images of n types of parts of the upper digestive tract as teacher data. As shown in Table 1 below, in this embodiment, n = 15.

[0039]

[0040] Each of the 15 types of parts corresponds to one of m types of organs. As shown in Table 2 below, in this embodiment, m = 6.

[0041]

[0042] The number of types of parts n is, for example, 2 or more and 100 or less. The number of types of organs m is (m < n) and is, for example, 2 or more and 20 or less. And the lesion detection unit 20 has m types of organ CADs corresponding to the m types of organs.

[0043] Note that the part inference unit 13 may be configured to infer (n + 2) part confidence levels for each of the endoscopic image being an image of n types of parts, an inappropriate image, and an undecidable image, using one or more models.

[0044] <Step S30> The part identification part 14 identifies the endoscopic image with the maximum part confidence RP as the image of part X. In other words, part X is identified as the candidate for the observed part during observation.

[0045] <Step S40> The first organ identification part 15 identifies the organ Y corresponding to part X as the first observed organ.

[0046] <Step S50> While the determination device 10 of the organ confidence inference processor 2 is performing the process from Step 20, it also performs the process from Step S50. Note that the process from Step S50 may be performed after the process from Step 20, or the process from Step 20 may be performed after the process from Step S50.

[0047] The organ inference unit 17 uses one second model learned with a plurality of images of a plurality of organs as teacher data to infer one second organ candidate corresponding to each endoscopic image.

[0048] That is, the organ inference unit 17 infers the organ confidence OPM (M = 1 to m) indicating that the input endoscopic image is an image of each of the m plurality of organs.

[0049] <Step S60> The organ identification unit 15 of the second organ candidate identification processor 2 identifies, for example, the organ with the highest organ confidence OP as the second organ candidate.

[0050] <Step S70> Observed organ determined? When the first organ candidate and the second organ candidate are the same, the organ determination unit 18 determines the first organ candidate (the same as the second organ candidate) as the observed organ in the endoscopic image (S70: YES). Then, in Step S80, it is determined whether the third model corresponding to the determined observed organ has already been activated. <G

[0051] When the first organ candidate and the second organ candidate are different (S70: NO), the determination device 10 does not determine the observed organ, and the lesion detection process is performed using the third model that has already been activated in Step S100.

[0052] Note that the organ inference unit 17 may infer a plurality of second organ candidates. For example, the organ inference unit 17 may infer two organs with an organ confidence OP equal to or higher than the first determination value TP1 as the second organ candidates (second organ candidate 1 and second organ candidate 2), respectively. If the organ confidence OP is less than the first determination value TP1, it is inferred that the image is not an image of the organ.

[0053] If multiple candidates for a second organ are inferred, the organ determination unit 18 determines the first organ candidate (which is the same as the second organ candidate) as the organ observed in the endoscopic image if any of the multiple candidates for a second organ is the same as the first organ candidate.

[0054] If all of the multiple organ confidence scores OP inferred by the organ inference unit 17 are below a predetermined organ confidence threshold, it is acceptable for none of the organs to be inferred as the second organ candidate. In this case, the organ determination unit 18 does not determine the observed organ.

[0055] <Step S80> Is the corresponding third model running? The CPU 11, for example, determines whether the third model (organ CAD) corresponding to the confirmed observed organ is running. If the third model (organ CAD) corresponding to the observed organ is running (YES), the process proceeds to step S100. In other words, the running third model continues to be used.

[0056] <Step S90> Activation of the corresponding second model (ON) The ON / OFF control unit 29 of the lesion detection unit 20 activates the CAD (third model) corresponding to the observed organ. The ON / OFF control unit 29 also stops (OFFs) all CADs other than the CAD (third model) corresponding to the observed organ.

[0057] <Step S100> Lesion detection processing The activated CAD of the lesion detection unit 20 infers the lesion from the endoscopic image. For example, the activated gastric CAD 30 performs inference using the second model 5. The inference result is output as a confidence level PP (PP = 0 to 1) inferred by the CAD corresponding to the observed organ.

[0058] For example, the confidence level PP1 for normal tissue is 0.66, and the confidence level PP2 for diseased tissue is 0.34 (PP1 + PP2 = 1.0). The lesion detection unit 20 may output either confidence level PP1 or confidence level PP2.

[0059] Furthermore, the "organ CAD" may include a "boundary CAD" that performs lesion detection processing on images containing multiple organs. For example, if the confidence level of site 4 (esophagus) and the confidence level of site 5 (esophagogastric junction) are both above a predetermined confidence level P, the boundary organ including the esophagus and stomach may be identified, and the corresponding boundary CAD may infer lesions in the endoscopic image.

[0060] Furthermore, when processor 2 starts up, i.e., at the start of the examination, there are no organs identified as being observed, and therefore no CAD is running. For this reason, lesion detection processing is not performed.

[0061] <Step S110> The display device 30 displays the observation site identified by the site identification unit 14, and / or the observation organ identified by the organ identification unit 15, as numerical values, illustrations, colors, etc. The display device 30 may also serve as a notification unit that notifies the observed organ.

[0062] Figure 3 shows an example of the screen of the display device 30. The screen in Figure 3 displays the endoscopic image 30A, the status of the judgment device 10 (AUTO mode) 30B, and the name of the detected observation site (antrum) as "Detected Site" 30C. Furthermore, the fact that the CAD currently in operation is (stomach) is graphically displayed using a figure 30D representing a stomach.

[0063] Although not shown in the diagram, if the confidence level PP2 for identifying lesion tissue is above a predetermined threshold, the lesion candidate may be superimposed on the endoscopic image 35A, or the confidence level PP2 may be displayed numerically or in color. For example, lesion candidates with a confidence level PP2 of 0.7 or higher may be superimposed on the endoscopic image 35A in red. Conversely, if the confidence level PP2 for identifying lesion tissue is below the predetermined threshold, no display is performed by the lesion detection unit 20. Then, the process from step S10 is repeated.

[0064] The operation method of processor 2 will be specifically explained using Table 3 shown below. Note that in Table 3, organs where the organ confidence level OP is less than the first judgment value TP1 are not shown.

[0065] 3

[0066] In Case 1, both the first and second organ candidates are the same, the "pharynx." Therefore, the "pharynx" is determined as the organ to be observed. In Case 2, the first and second organ candidates are different. Therefore, no new organ to be observed is determined, and the operating CAD system infers the lesion. In Case 3, the first organ candidate and the second organ candidate 1 are different. However, the first organ candidate and the second organ candidate 2 are the same. Therefore, the "stomach" is determined as the organ to be observed.

[0067] As described above, the operation method of the endoscopic image diagnostic support processor of this embodiment is as follows: For multiple endoscopic images that are input in succession, a first model that has been trained using multiple images of multiple body parts as training data is used to infer the confidence level of each body part that each endoscopic image is an image of one of the multiple body parts; based on predetermined conditions using the confidence level of each body part, the observation area of ​​the endoscopic image is identified and a first organ candidate corresponding to the observation area is identified; at least one second organ candidate corresponding to each endoscopic image is inferred using at least one second model that has been trained using multiple images of at least one organ as training data; if the first organ candidate and the second organ candidate are the same, the first organ candidate is determined as the observation organ of the endoscopic image; if the first organ candidate and the second organ candidate are different, the observation organ is not determined, or the observation organ is determined based on predetermined conditions.

[0068] The program for the endoscopic image diagnostic support processor according to an embodiment of the present invention, for a plurality of endoscopic images that are input sequentially, uses a first model that has been trained using multiple images of multiple body parts as training data to infer the confidence level of each body part that each endoscopic image is an image of one of the multiple body parts, identifies the observation site of the endoscopic image and identifies a first organ candidate corresponding to the observation site based on predetermined conditions using the confidence level of each body part, uses at least one second model that has been trained using multiple images of at least one organ as training data to infer at least one second organ candidate corresponding to each endoscopic image, if the first organ candidate and the second organ candidate are the same, confirms the first organ candidate as the observation site of the endoscopic image, and if the first organ candidate and the second organ candidate are different, either does not confirm the observation site or causes the computer to execute a process to confirm the observation site based on predetermined conditions.

[0069] The non-temporary storage medium of the embodiment of the present invention stores the above program.

[0070] The non-temporary storage medium, the storage device 32, may store only a part of the program. Furthermore, the program may be distributed or provided via a communication network. Users can execute all or part of the processing, and thus perform the processing of the processor described above, by installing the program from the storage medium to their computer, or by downloading the program via the communication network and installing it to their computer.

[0071] <Second Embodiment> The endoscopic image diagnostic support processor 2A (processor 2A) of this embodiment, described below, is similar to the endoscopic image diagnostic support processor 2 of the first embodiment and has the same effects as the endoscopic image diagnostic support processor 2. For this reason, components with the same functions as the endoscopic image diagnostic support processor 2 are denoted by the same reference numerals as the endoscopic image diagnostic support processor 2, and their descriptions are omitted.

[0072] As shown in Figure 4, in the processor 2A, each of the multiple third models 21 to 26 of the lesion detection unit 20A, which have been trained using multiple lesion images and multiple normal images of their respective corresponding organs as training data, is considered to include an organ inference unit 1 to 6 (17A to 17F), which is their respective second model. That is, each of the organ inference units 1 to 6 (17A to 17F) uses a multiple second model, each trained using multiple images of their respective corresponding organs as training data, to infer multiple second organ candidates corresponding to the endoscopic image.

[0073] The organ inference units 17A to 17F infer whether the endoscopic image corresponds to the corresponding organ that they have learned.

[0074] The flowchart in Figure 5 shows how processor 2A operates. In step S55, in processor 2A, each of the multiple organ inference units 17A to 17F (multiple second models) infers an organ confidence score OP, which indicates that the endoscopic image is an image of its corresponding organ. For example, organ inference unit 17 infers that the endoscopic image corresponds to the corresponding organ if the organ confidence score OP is greater than or equal to a predetermined value, and infers that the endoscopic image does not correspond to the corresponding organ if the organ confidence score OP is less than the predetermined value.

[0075] In step S70, if the first organ candidate and the second organ candidate are the same, the organ determination unit 18 determines the first organ candidate as the organ to be observed in the endoscopic image (YES). If the first organ candidate and the second organ candidate are different, the organ determination unit 18 either does not determine the organ to be observed (NO), or determines the organ to be observed based on predetermined conditions.

[0076] The operation method of processor 2A will be explained in detail using Table 4 below. Note that, for the sake of simplicity, only the "pharynx," "esophagus," and "stomach" are shown in Table 4.

[0077]

[0078] In Case 1, for the first organ candidate, "pharynx," among the multiple second models, the pharyngeal CAD estimates it to be "pharynx," while the other CADs estimate it to be not the corresponding organ. Since the estimation results of the first organ candidate and the multiple second models match, the confirmed organ is "pharynx."

[0079] In Case 2, all of the second models estimate that the first organ candidate, "duodenum," is not the corresponding organ. Therefore, since the estimation results of the first organ candidate and the multiple second models match, the confirmed organ is "duodenum." If the duodenum CAD infers that it is "not the duodenum," the organ confirmation unit 18 does not confirm the observed organ.

[0080] In Case 3, for the first organ candidate, "pharynx," the pharyngeal CAD estimated "pharynx," the esophageal CAD estimated "esophagus," and the stomach CAD estimated "stomach." Since the answers from the first organ candidate and the multiple second models match, the confirmed organ is "pharynx."

[0081] In cases 4 and 5, the organ determination unit 18 does not determine the observed organ because the estimation results of multiple second models do not match for the first organ candidate, "pharynx."

[0082] Processor 2A can identify the observed organ more efficiently than processor 2.

[0083] The range of values ​​described above is not limited to that range and can be increased or decreased as appropriate. Furthermore, the present invention is not limited to the embodiments described above, and various changes and modifications can be made as long as they do not alter the essence of the invention.

[0084] 1... Endoscopy system 2, 2A... Endoscopy image diagnostic support processor (processor) 9... Endoscope 10... Observed organ determination device (determination device) 11... CPU 12... Image input unit 13... Site inference unit 14... Site identification unit 15... Organ identification unit 16... Memory 17... Organ inference unit 18... Organ confirmation unit 20... Lesion detection unit 30... Display device 31... Notification unit 32... Storage device 40... Server

Claims

1. An endoscopic image diagnostic support processor comprising: a site inference unit that, for multiple endoscopic images input in succession, uses a first model trained on multiple images of multiple sites as training data to infer the site confidence level for each endoscopic image that is an image of one of the multiple sites; a site identification unit that identifies the observation site of the endoscopic image based on predetermined conditions using the site confidence level; an organ identification unit that identifies a first organ candidate corresponding to the observation site; an organ inference unit that infers at least one second organ candidate corresponding to each of the endoscopic images using at least one second model trained on multiple images of at least one organ as training data; and an organ identification unit that, if the first organ candidate and the second organ candidate are the same, determines the first organ candidate as the observation organ of the endoscopic image, and if the first organ candidate and the second organ candidate are different, either does not determine the observation organ or determines the observation organ based on predetermined conditions.

2. The endoscope image diagnostic support processor according to claim 1, wherein the site identification unit identifies the site with the highest confidence level as the observation site.

3. The endoscopic image diagnostic support processor according to claim 1, wherein the organ inference unit infers the organ confidence level for each of the plurality of organs, and infers the at least one second organ candidate based on the organ confidence level.

4. The endoscopic image diagnostic support processor according to claim 3, wherein the organ inference unit infers the organ with the highest organ confidence level as the second organ candidate.

5. The endoscopic image diagnostic support processor according to claim 3, wherein the organ inference unit infers each of a plurality of organs whose organ confidence level is equal to or greater than a first determination value as a second organ candidate, and the organ determination unit determines the first organ candidate as the observed organ in the endoscopic image if any of the plurality of second organ candidates is the same as the first organ candidate, and does not determine the observed organ or determines the observed organ based on predetermined conditions if all of the plurality of second organ candidates are different from the first organ candidate.

6. The endoscopic image diagnostic support processor according to claim 1, comprising a notification unit that notifies at least one of the observation site and the observation organ.

7. The endoscopic image diagnostic support processor according to claim 1, further comprising a lesion detection unit for inferring lesions in the endoscopic image, wherein the lesion detection unit infers the lesion using a third model corresponding to the observed organ determined by the organ determination unit, from among a plurality of third models which have been trained using a plurality of lesion images and a plurality of normal images as training data for a plurality of organs.

8. If the organ determination unit does not determine the organ to be observed, the lesion detection unit uses the operating third model, as described in claim 7, for the endoscopic image diagnostic support processor.

9. The endoscopic image diagnostic support processor according to claim 1, wherein each of a plurality of third models, each trained using multiple lesion images and multiple normal images for each of multiple organs as training data, includes a second model that infers whether the endoscopic image corresponds to the corresponding organ that each third model has trained.

10. A method for operating an endoscopic image diagnostic support processor, which, for multiple endoscopic images input in succession, uses a first model trained on multiple images of multiple body parts as training data to infer a confidence level for each endoscopic image that each of the multiple body parts is one of the images of the multiple body parts; identifies the observation site of the endoscopic image and identifies a first organ candidate corresponding to the observation site based on predetermined conditions using the confidence level; uses at least one second model trained on multiple images of at least one organ as training data to infer at least one second organ candidate corresponding to each of the endoscopic images; if the first organ candidate and the second organ candidate are the same, confirms the first organ candidate as the observation organ of the endoscopic image; and if the first organ candidate and the second organ candidate are different, either does not confirm the observation organ or confirms the observation organ based on predetermined conditions.

11. A program for an endoscopic image diagnostic support processor that, for multiple endoscopic images input in succession, uses a first model trained on multiple images of multiple body parts as training data to infer a confidence level for each endoscopic image that each of the multiple body parts is one of the images of the multiple body parts; identifies the observation site of the endoscopic image and identifies a first organ candidate corresponding to the observation site based on predetermined conditions using the confidence level; uses at least one second model trained on multiple images of at least one organ as training data to infer at least one second organ candidate corresponding to each of the endoscopic images; if the first organ candidate and the second organ candidate are the same, confirms the first organ candidate as the observation site of the endoscopic image; and if the first organ candidate and the second organ candidate are different, either does not confirm the observation site or causes the computer to perform a process to confirm the observation site based on predetermined conditions.

Images