Pathological evaluation method and analysis device for endoscopic submucosal dissection

Abstract

A pathological evaluation method for endoscopic submucosal dissection comprises the steps in which: an analysis device receives an input of a Haematoxylin and eosin (H&E) stained image of a specimen; the analysis device identifies a region of interest by inputting the H&E stained image to a segmentation model; and the analysis device evaluates a submucosal invasion state of the specimen according to the ratio of a tumor region in a specific muscularis mucosae region of the region of interest.

Description

Pathological evaluation method and analysis device for endoscopic submucosal resection

[0001] The technology described below relates to a technique for predicting resection for the treatment of gastric cancer patients using a deep learning model.

[0002] Gastric cancer is a tumor with a high incidence rate worldwide. In particular, the proportion of early-stage gastric cancer is on the rise. Endoscopic submucosal dissection (ESD) is the standard treatment for early-stage gastric cancer with a low risk of lymph node metastasis. Specimens obtained during the ESD procedure are crucial for determining whether further treatment is necessary. If the pathological diagnosis of the specimen indicates a risk of lymph node metastasis, a gastrectomy including lymph node dissection is performed on the patient.

[0003] Resection for gastric cancer patients is determined based on the results of a tissue biopsy. Generally, medical professionals determine the type or extent of resection based on tissue staining results, such as H&E (Haematoxylin & eosin) staining.

[0004] The technology described below aims to provide a technique capable of automatically evaluating whether to perform endoscopic submucosal resection based on a patient's H&E stained image.

[0005] A pathological evaluation method for endoscopic submucosal resection includes the steps of: an analysis device receiving an H&E stained image of a sample; the analysis device inputting the H&E stained image into a segmentation model to distinguish a region of interest; and the analysis device evaluating the submucosal infiltration status of the sample based on the ratio of a tumor region to a specific mucosal muscle region of the region of interest.

[0006] An analysis device for evaluating the state of submucosal infiltration includes an interface device for receiving an H&E stained image of a sample, a storage device for storing a segmentation model that distinguishes a region of interest in a tissue stained image, and a computing device that inputs the received H&E stained image into the segmentation model to distinguish the region of interest and evaluates the state of submucosal infiltration of the sample according to the ratio of a tumor region to a specific mucosal muscle region of the region of interest.

[0007] The technology described below can precisely evaluate the state of gastric cancer invasion by utilizing an artificial intelligence model that interprets H&E stained images. By quantitatively evaluating the degree of tumor invasion in the region of interest, the technology can provide diagnostic support information, such as recommendations for optimal resection.

[0008] Figure 1 is an example of a system that recommends resection for a tumor using tissue staining images.

[0009] Figure 2 is an example of training data used for training a segmentation model.

[0010] Figure 3 is an example of the learning process of a segmentation model.

[0011] Figure 4 is an example of a process for evaluating the state of submucosal infiltration by detecting tumor areas and submucosal areas.

[0012] Figure 5 shows the results of comparing the segmentation results.

[0013] Figure 6 shows the results of evaluating the performance of the segmentation model.

[0014] Figure 7 is an example of an analysis device for evaluating resection surgery on a gastric cancer patient.

[0015] The technology described below is subject to various modifications and may have various embodiments, and specific embodiments are illustrated in the drawings and described in detail. However, this is not intended to limit the technology described below to specific embodiments, and it should be understood that it includes all modifications, equivalents, and substitutions that fall within the spirit and scope of the technology described below.

[0016] Terms such as first, second, A, B, etc., may be used to describe various components, but such components are not limited by the said terms and are used solely for the purpose of distinguishing one component from another. For example, without departing from the scope of rights of the technology described below, the first component may be named the second component, and similarly, the second component may be named the first component. The term "and / or" includes a combination of multiple related described items or any of the multiple related described items.

[0017] In terms used in this specification, singular expressions should be understood to include plural expressions unless the context clearly indicates otherwise, and terms such as “includes” should be understood to mean that the described features, number, steps, actions, components, parts, or combinations thereof exist, and not to exclude the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0018] Before providing a detailed description of the drawings, it is to clarify that the classification of components in this specification is merely based on the primary function each component is responsible for. That is, two or more components described below may be combined into a single component, or a single component may be divided into two or more components based on more subdivided functions. Furthermore, each component described below may additionally perform some or all of the functions of other components in addition to its own primary function, and it goes without saying that some of the primary functions of each component may be exclusively performed by other components.

[0019] Furthermore, in performing the method or operation method, each process constituting the method may occur differently from the specified order unless a specific order is clearly indicated in the context. That is, each process may occur in the same order as specified, may be performed substantially simultaneously, or may be performed in the reverse order.

[0020] The technique described below is a method for evaluating the status of tumor infiltration within tissues based on tissue staining images of gastric cancer patients.

[0021] In addition, the technique described below is a method that provides diagnostic support information regarding the appropriate resection for a patient based on tissue staining images of gastric cancer patients. In this case, the resection may be an endoscopic submucosal resection.

[0022] The technique described below uses a deep learning model to identify tumor regions from H&E stained images.

[0023] Deep learning models can be segmentation models that distinguish regions of interest within an organization. For example, deep learning models can be semantic segmentation models such as U-Net, DeepLabv3, and SegFormer.

[0024] A device that determines the resection area or resection information using stained images of tissues is referred to as an analysis device. The analysis device is a device capable of data processing and may take the form of a PC, smart device, server, etc.

[0025] FIG. 1 is an example of a system (100) that recommends resection of a tumor using a tissue staining image. In FIG. 1, the analysis device is illustrated as an example of a computer terminal (130) and a server (150).

[0026] The staining image generating device (110) is a device that stains a tissue slide and scans the stained result to generate a stained image. The staining image generating device (110) can generate an H&E stained image of a sample. At this time, the tissue may be the above tissue.

[0027] H&E staining stains tissue samples with hematoxylin and eosin. H&E staining allows for the clear differentiation of cellular structures such as the cytoplasm, nucleus, organelles, and extracellular matrix. An H&E stained image is a digital (scanned) image of a stained tissue slide.

[0028] The computer terminal (130) is a device used by user A for organizational analysis.

[0029] The computer terminal (130) can receive an H&E dyeing image from the dyeing image generating device (110) via a wired or wireless network.

[0030] The computer terminal (130) can distinguish regions of interest based on the H&E stained image. At this time, the regions of interest may include tumor regions and mucosal muscle regions. The computer terminal (130) can distinguish regions of interest in the H&E stained image using a segmentation model. The process of constructing the segmentation model will be described later.

[0031] The computer terminal (130) can determine a suitable resection for the sample based on the infiltration status (submucosal infiltration) or degree of infiltration of the region of interest in the sample tissue. In this case, the resection may be an endoscopic submucosal resection (ESD).

[0032] User A can perform a clinical evaluation of the sample through the computer terminal (130). User A can check the tumor infiltration status (submucosal infiltration) or degree of infiltration of the sample slide through the computer terminal (130). Additionally, User A can check information on resection suitable for the sample (patient) through the computer terminal (130).

[0033] FIG. 1 illustrates a dyeing image generating device (110) and a computer terminal (130) as separate objects. In some cases, the dyeing image generating device (110) and the computer terminal (130) may be physically implemented as a single device or connected devices.

[0034] The server (150) can receive an H&E dyeing image from the dyeing image generating device (110) via a wireless network.

[0035] The server (150) can distinguish regions of interest based on H&E stained images. At this time, the regions of interest may include tumor regions and mucosal muscle regions. The server (150) can distinguish regions of interest in H&E stained images using a segmentation model. The process of constructing the segmentation model will be described later.

[0036] The server (150) can determine a suitable resection for a sample based on the infiltration status (submucosal infiltration) or degree of infiltration of the region of interest in the sample tissue. In this case, the resection may be an endoscopic submucosal resection (ESD).

[0037] The server (150) can transmit the analysis results to the user terminal (50). User A can perform a clinical evaluation of the sample through the user terminal (50). User A can check the tumor infiltration status (submucosal infiltration) or degree of infiltration of the sample slide through the user terminal (50). Additionally, User A can check information on the appropriate resection procedure for the sample (patient) through the user terminal (50).

[0038] Meanwhile, the computer terminal (130) or server (150) can store the analysis results in a separate device such as an EMR (Electronic Medical Record).

[0039] The region of interest may include the tumor region and the muscularis mucosa region. Alternatively, the region of interest may include the tumor region, the muscularis mucosa region, and the remaining other regions. The segmentation model can distinguish the tumor region and the muscularis mucosa region in the tissue. Alternatively, the segmentation model can distinguish the tumor region, the muscularis mucosa region, and the remaining other regions. Figure 2 is an example of training data used to train a segmentation model. Figures 2(A), 2(B), and 2(C) are gastric tissue slides.

[0040] FIG. 3 is an example of the learning process (200) of a segmentation model. The segmentation model can be built through a separate learning device. In this case, the learning device may be a device separate from the analysis device. The learning device is a computing device capable of performing data processing, image processing, deep learning model learning, etc.

[0041] The learning device constructs the learning data (210).

[0042] The training database (DB) stores the training data set.

[0043] The training dataset includes "pairs of H&E stained images and mask images with regions of interest marked (annotated) on the stained images" for multiple samples. The mask with the region of interest marked is named a label mask. In Figure 2, the training dataset illustrates H&E stained images and label masks for n samples.

[0044] The label mask may be an image in which a professional has manually masked the region of interest on the original H&E stained image. The label mask may be an image divided into tumor region, mucomuscular region, and other regions.

[0045] The learning device builds a segmentation model using the learning data (220).

[0046] The learning device inputs the H&E stained images (input images) of the training data into a segmentation model. The segmentation model may be a semantic segmentation model such as a U-net. The segmentation model can receive H&E stained images as input and output three channel probability maps. The three channel probability maps correspond to the tumor region, the mucomuscular region, and other regions, respectively. The channel probability maps can represent their probabilities using color.

[0047] The learning device compares the input three-channel probability map and label mask to update the parameters of the segmentation model so that the segmentation model can accurately distinguish the region of interest.

[0048] The learning device can repeat the learning process of the segmentation model using multiple training data.

[0049] A segmentation model was constructed using the dataset, and its performance was evaluated. The dataset was obtained from ESD specimens of gastric cancer patients who underwent endoscopic submucosal dissection (ESD) at Samsung Seoul Hospital between January 2020 and June 2024. The ESD specimens were sliced ​​at 2 mm intervals. The average number of slides per patient varied from 3 to 26, and the number of sections per slide was generally 3.

[0050] During the model training process, representative slides selected for each patient were used as training data, while either the patient's representative slides or all slides were used as validation data. The dataset included only samples histologically diagnosed as adenocarcinoma, and samples diagnosed as other tumor types, such as precancerous lesions or neuroendocrine tumors, were excluded.

[0051] The clinical pathology information for the entire dataset is shown in Table 1 below. The dataset was divided into a development dataset, a test dataset for submucosal infiltration detection tests, and a dataset for pathological performance evaluation. For the performance evaluation dataset, 10 samples were selected from the test dataset.

[0052] Item Segmentation model training dataset (n = 140) Test dataset for WSI level submucosal invasion detection (n = 61) Dataset for pathological performance evaluation (n = 10) Gender Male 97 (69.3%) 43 (70.5%) 8 (80%) Female 43 (30.7%) 18 (29.5%) 2 (20%) Age 64.6 years (Standard deviation, SD 9.7) 69.1 years (SD 8.0) 70.8 years (SD 6.4) Lesion size (mm) 17.1 (SD 9.8) 18.6 (SD 9.1) 14 (SD 5.6) History type according to WHO classification Well-differentiated tubular adenocarcinoma 48 (34.3%) 12 (19.7%) 4 (40%) Moderately differentiated Tubular adenocarcinoma (moderately differentiated) 74 (52.9%) 41 (67.2%) 6 (60%) Poorly differentiated tubular adenocarcinoma 1 (0.7%) 1 (1.6%) 0 (0%) Poorly cohesive carcinoma 14 (10.0%) 1 (1.6%) 0 (0%) Other 3 (2.1%) 6 (9.8%) 0 (0%) Presence of undifferentiated component (≥ 5%) Present 26 (18.6%) 10 (16.4%) 0 (0%) None 114 (81.4%) 51 (83.6%) 10 (100%) Intestinal Lauren classification type 115 (82.1%) 56 (91.8%) 10 (100%) Diffuse type 11 people (7.9%) 1 person (1.6%) 0 people (0%) Mixed type 12 people (8.6%) 1 person (1.6%) 0 people (0%) Indeterminate type 2 people (1.4%) 3 people (4.9%) 0 people (0%) Depth of infiltration Mucosa 95 people (67.9%) 18 people (29.5%) 3 people (30%) Submucosal 45 people (32.1%) 43 people (70.5%) 7 people (70%) Lymphatic invasion present 23 people (16.4%) 23 people (37.7%) 2 people (20%) No 117 people (83.6%) 38 people (62.3%) 8 people (80%) Resection margin No tumor 130 people (92.9%) 52 people (85.2%) 9 people (90%) Tumor invasion present 10 people (7.1%) 9 people (14.8%) 1 person (10%).

[0053] Segmentation models were constructed using U-Net++, DeepLabv3+, and SegFormer, respectively. U-Net++ and DeepLabv3+ are convolutional neural network-based models. SegFormer is a transformer-based model. Transformer-based models can process input images by dividing them into multiple patches.

[0054] The segmentation model was constructed using a loss function as shown in Equation 1 below. The loss function is set to minimize the difference between the label mask and the region of interest distinguished by the model.

[0055]

[0056] In Equation 1, H and W are the image height and width, and C is the number of classes. y ijc is binary information (0 or 1) indicating whether an accurate classification was performed at pixel (i,j). is the probability value of class c belonging to pixel (i,j).

[0057] Figure 4 is an example of a process (300) for detecting tumor areas and submucosal areas to evaluate the state of submucosal infiltration.

[0058] The analysis device receives an H&E stained image (310). At this time, the analysis device can process the H&E stained image by dividing it into multiple patches. In this case, the analysis device can adjust the patch size to match the image size used during the training process of the segmentation model. That is, if the segmentation model is trained in units of patches of a certain size during the training process, the analysis device can divide the H&E stained image into multiple patches and distinguish regions of interest for each patch.

[0059] The analysis device can input the input H&E stained image into a segmentation model to distinguish regions of interest (320). If the segmentation model processes the image in patch units, the analysis device can stitch the patches together to form a single slide-sized image. The image generated by the segmentation model can distinguish between tumor regions and mucosal muscle regions.

[0060] The analysis device can evaluate the tumor infiltration status in the mucosal muscle region (330). The analysis device can evaluate the tumor infiltration status in the mucosal muscle region using a grid having multiple square regions or points. FIG. 4 illustrates an example of setting up a grid in three mucosal muscle regions (a, b, and c).

[0061] The grid can be composed of n×n square regions (points). The analysis device can evaluate the tumor invasion status for a single grid set in the mucosal region. The analysis device identifies the number of regions m where the tumor is located among the square regions belonging to a single grid. If the analysis device determines that there are many tumor regions among the square regions belonging to the grid, it can determine that region as submucosal invasion. The analysis device can determine submucosal invasion based on the number or ratio of tumor regions belonging to the grid. The analysis device can calculate the probability p of submucosal invasion based on the maximum value as shown in Equation 2 below.

[0062]

[0063] m in mathematical formula 2 max is the maximum number of sub-regions that are tumors belonging to one of the grids set in the region of interest, and c is the threshold.

[0064] The analysis device may also calculate the submucosal invasion probability p based on the mean value as shown in Equation 3 below. The mean value-based method is an example of calculating the submucosal invasion probability p based on multiple grids set on a single slide.

[0065]

[0066] m in mathematical formula 3 i n is the number of square regions that are tumors in grid i, and k is m among all grids. i It is the number of grids that are >c.

[0067] The analysis device can evaluate the submucosal infiltration status of the current sample slide based on the infiltration probability p. For example, the analysis device can determine that the sample has submucosal infiltration if the infiltration probability p is above a certain threshold.

[0068] The analysis device can superimpose images of regions of interest (tumor region and mucosal muscle region) distinguished by the segmentation model onto H&E stained images (340). Through this, the analysis device can visualize the regions of interest within the tissue.

[0069] Figure 5 shows the results of comparing the segmentation results. Figure 5 shows the results of validating the segmentation model using the validation dataset. In Figure 5, the original image is the original H&E stained image, the annotated image is the image with the region of interest manually labeled (correct answer), and the predicted image is the image in which the segmentation model has distinguished the region of interest. Looking at Figure 5, it can be seen that the segmentation model distinguishes the region of interest close to the correct answer.

[0070] Figure 6 shows the results of evaluating the performance of the segmentation model. Figure 6 shows the results of verifying the performance of the segmentation model using a validation dataset.

[0071] Figure 6(A) shows the performance of the segmentation model in distinguishing tumor regions. The segmentation model showed a performance of AUC (Area Under the ROC Curve) = 0.9951.

[0072] Figure 6(B) shows the performance of classifying regions of interest using a segmentation model and evaluating submucosal infiltration based on the maximum value (Equation 2). The evaluation of submucosal infiltration showed performance exceeding an AUC of 0.9. The grid size was verified by setting it to various values.

[0073] Figure 6(C) shows the performance of classifying regions of interest using a segmentation model and evaluating submucosal infiltration based on the mean value (Equation 3). The evaluation of submucosal infiltration showed performance exceeding an AUC of 0.9. The grid size was verified by setting it to various values.

[0074] FIG. 7 is an example of an analysis device (400) for evaluating resection surgery on a gastric cancer patient. The analysis device (400) corresponds to the analysis device described above (130 and 150 in FIG. 1). The analysis device (400) can be physically implemented in various forms. For example, the analysis device (400) can take the form of a computer device such as a PC, a server of a network, a chipset dedicated to data processing, etc.

[0075] The analysis device (400) may include a storage device (410), a memory (420), a computation device (430), an interface device (440), a communication device (450), and an output device (460).

[0076] The storage device (410) can store the H&E stained image of the sample.

[0077] The storage device (410) can store the aforementioned segmentation model. The segmentation model is a pre-trained model. The segmentation model can generate an image in which regions of interest in an input image are distinguished by color.

[0078] The storage device (410) can store an image in which a region of interest is separated from an H&E stained image.

[0079] The storage device (410) can store the result of predicting submucosal infiltration using the result of distinguishing the region of interest using a segmentation model.

[0080] The memory (420) can store data and information generated during the process of the analysis device evaluating the resection of a gastric cancer patient.

[0081] The interface device (440) is a device that receives certain commands and data from the outside.

[0082] The interface device (240) may be configured to receive user commands. Alternatively, the interface device (240) may be configured to receive certain data from an external storage medium. Alternatively, the interface device (240) may be configured to receive data received through a communication device (250).

[0083] The interface device (440) can receive the patient's H&E staining image from a physically connected input device or an external storage device.

[0084] The interface device (440) may also transmit an image in which a region of interest is separated from an H&E stained image to an external object.

[0085] The interface device (440) may also transmit the result of predicting submucosal infiltration to an external object.

[0086] The communication device (450) means a configuration that receives and transmits certain information through a wired or wireless network.

[0087] The communication device (450) can receive the patient's H&E staining image from an external object.

[0088] The communication device (450) may also transmit an image in which a region of interest is identified in an H&E dyed image to an external object such as a user terminal.

[0089] The communication device (450) may also transmit the result of predicting submucosal infiltration to an external object such as a user terminal.

[0090] The output device (460) is a device that outputs certain information.

[0091] The output device (460) can output interfaces required for the analysis process, results of distinguishing regions of interest in H&E stained images, and results of submucosal infiltration prediction.

[0092] The computing device (430) can preprocess the H&E dyed image in a consistent manner. For example, the computing device (430) can normalize the H&E dyed image in a consistent manner. The computing device (430) can adjust the resolution of the H&E dyed image. In addition, the computing device (430) can normalize the color of the H&E dyed image.

[0093] The computing device (430) can divide the H&E dyed image into multiple patches.

[0094] The computing device (430) can input the H&E stained image into a segmentation model to distinguish regions of interest.

[0095] The computing device (430) can input each patch of the H&E stained image into a segmentation model to distinguish regions of interest on a patch-by-patch basis. In this case, the computing device (430) can combine the results of distinguishing regions of interest for multiple patches to generate a single slide-sized image.

[0096] The computing device (430) can evaluate whether there is submucosal infiltration using the result of distinguishing the region of interest in the segmentation model.

[0097] The computing device (430) can set an n×n grid in the mucosal muscle region of the distinguished region of interest and check the number of tumor sub-regions in the grid.

[0098] The computing device (430) can evaluate the submucosal infiltration status based on the maximum number of tumor subregions in any one of the grids (see Equation 2).

[0099] The computing device (430) can evaluate the submucosal infiltration status based on the case where the number of tumor subregions in each of the multiple grids is greater than or equal to a threshold value (see Equation 3).

[0100] That is, the computing device (430) can evaluate the submucosal infiltration status according to the ratio of the tumor area to the mucosal muscle area of ​​the tissue slide.

[0101] If the computing device (430) evaluates the region of interest as submucosal infiltration, it can determine the resection for the sample as endoscopic submucosal resection.

[0102] The computing device (430) may be a device such as a processor, AP, or a chip with a program embedded in it that processes data and performs certain operations.

[0103] The methods according to the embodiments described in the specification of the present disclosure may be implemented in the form of hardware, software, or a combination of hardware and software.

[0104] When implemented in software, a computer-readable storage medium may be provided for storing one or more programs (software modules). One or more programs stored in the computer-readable storage medium are configured for execution by one or more processors within an electronic device. One or more programs include instructions that cause the electronic device to execute methods according to embodiments described in the specification of this disclosure.

[0105] In addition, the staining image analysis method and the method for providing treatment information for gastric cancer patients described above may be implemented as a program (or application) comprising an executable algorithm that can be executed on a computer. The program may be provided by storing it on a transitory or non-transitory computer-readable medium.

[0106] A non-transient readable medium refers to a medium that stores data semi-permanently and can be read by a device, rather than a medium that stores data for a short period of time, such as a register, cache, or memory. Specifically, the various applications or programs described above may be stored and provided on a non-transient readable medium such as a CD, DVD, hard disk, Blu-ray disc, USB, memory card, ROM (read-only memory), PROM (programmable read-only memory), EPROM (Erasable PROM, EPROM), EEPROM (Electrically EPROM), or flash memory.

[0107] Transient readable media refers to various types of RAM such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDR SDRAM), Enhanced SDRAM (ESDRAM), Synclink DRAM (SLDRAM), and Direct Rambus RAM (DRRAM).

[0108] Additionally, the program may be stored on an attachable storage device that can be accessed via a communication network such as the Internet, Intranet, LAN (local area network), WAN (wide area network), or SAN (storage area network), or a combination thereof. Such a storage device may be connected to a device performing an embodiment of the present disclosure through an external port. Additionally, a separate storage device on a communication network may be connected to a device performing an embodiment of the present disclosure.

[0109] The embodiments and drawings attached to this specification merely clearly illustrate a part of the technical ideas included in the aforementioned technology, and it is self-evident that all variations and specific embodiments that can be easily inferred by a person skilled in the art within the scope of the technical ideas included in the specification and drawings of the aforementioned technology are included within the scope of the rights of the aforementioned technology.

Claims

1. A step in which the analysis device receives an H&E (Haematoxylin and eosin) stained image of the sample; The step of the analysis device inputting the H&E stained image into a segmentation model to distinguish regions of interest; and The analysis device includes the step of evaluating the submucosal infiltration status of the sample according to the ratio of the tumor area to a specific mucosal muscle area of ​​the area of ​​interest, and A method for pathological evaluation for endoscopic submucosal resection, wherein the above-mentioned region of interest includes a tumor region and a muscle mucosa region.

2. In Paragraph 1, A pathological evaluation method for endoscopic submucosal resection, wherein the above segmentation model is a model trained using training data comprising H&E stained images and a label mask annotated with the region of interest for the above H&E stained images.

3. In Paragraph 1, A pathological evaluation method for endoscopic submucosal resection, wherein the analysis device sets up n×n grids for a plurality of mucosal muscle regions located in the region of interest, and evaluates the sample as having a submucosal infiltration state when the maximum number of tumor sub-regions among the grids is greater than or equal to a threshold value.

4. In Paragraph 1, A pathological evaluation method for endoscopic submucosal resection, wherein the analysis device sets n×n grids for a plurality of mucosal muscle regions located in the region of interest, and evaluates the sample as having a submucosal infiltration state when the average value of the number of subregions that are tumorous among the grids is greater than or equal to a threshold value.

5. In Paragraph 1, A pathological evaluation method for endoscopic submucosal resection, comprising the step of determining the resection of the sample as endoscopic submucosal resection when the analysis device evaluates the sample as having a submucosal infiltration state.

6. Interface device for receiving an H&E (Haematoxylin and eosin) stained image of a sample; A storage device for storing a segmentation model that distinguishes regions of interest in a tissue staining image; and It includes a computing device that inputs the above-mentioned input H&E stained image into the above-mentioned segmentation model to distinguish regions of interest, and evaluates the submucosal infiltration status of the sample according to the ratio of the tumor region to a specific mucosal muscle region of the above-mentioned region of interest, and The above-mentioned region of interest is an analysis device for evaluating the state of submucosal infiltration, comprising a tumor region and a muscle mucosa region.

7. In Paragraph 6, An analysis device for evaluating submucosal infiltration status, wherein the above segmentation model is a model trained using training data including H&E stained images and a label mask in which the region of interest is annotated for the above H&E stained images.

8. In Paragraph 6, The above-described computing device sets n×n grids for a plurality of mucosal muscle regions located in the region of interest, and an analysis device for evaluating the submucosal infiltration state in which the sample is evaluated as a submucosal infiltration state when the maximum number of tumor sub-regions among the grids is greater than or equal to a threshold value.

9. In Paragraph 6, The above-described computing device sets up n×n grids for a plurality of mucosal muscle regions located in the region of interest, and an analysis device for evaluating the submucosal infiltration state in which the average value of the number of sub-regions that are tumorous among the grids and are greater than or equal to the threshold value evaluates the sample as a submucosal infiltration state.

10. In Paragraph 6, The above-mentioned computing device is an analysis device for evaluating the submucosal infiltration state, which determines the resection of the sample as endoscopic submucosal resection when the sample is evaluated as having a submucosal infiltration state.

Images