Method for screening normal tissue using artificial intelligence, and electronic device for performing same

Abstract

The present disclosure relates to a method for screening a normal tissue by using artificial intelligence, and an electronic device for performing same. The method according to one embodiment of the present disclosure may comprise the steps of: acquiring an entire image corresponding to a tissue sample; acquiring a plurality of region-of-interest images on the basis of the entire image; acquiring a plurality of patches on the basis of the plurality of region-of-interest images; extracting a plurality of signal values by using a first artificial intelligence model with the plurality of patches as input; acquiring a plurality of embedding vectors each corresponding to the plurality of region-of-interest images by using a second artificial intelligence model with the plurality of signal values and position values corresponding to the plurality of patches as input; determining whether a region-of-interest image having a lesion exists among the plurality of region-of-interest images by using a third artificial intelligence model with the plurality of embedding vectors as input; and determining that the tissue sample is normal on the basis of determining that no region-of-interest image having a lesion exists among the plurality of region-of-interest images.

Description

Method for screening normal tissue using artificial intelligence and electronic device for performing the same

[0001] The present disclosure relates to a method for screening normal tissue using artificial intelligence and an electronic device for performing the same, and more specifically, to a method and device capable of determining with high accuracy whether a tissue sample is normal by preprocessing an entire image corresponding to a tissue sample and using a plurality of artificial intelligence models to determine whether the tissue sample is normal tissue.

[0002] Advances in artificial intelligence and computer vision technologies are leading to active research into technologies that utilize these technologies to process medical data. In the medical field, there is a growing trend toward leveraging artificial intelligence to analyze pathological images faster and more accurately, moving beyond the traditional method of doctors manually analyzing pathological images. These technologies are significantly contributing to improving diagnostic accuracy and reducing time and costs in the field of pathology.

[0003] However, the resolution and size of pathological images pose a challenge for AI analysis, resulting in low accuracy. Pathological images consist of very high-resolution, large files, and processing them requires significant computational resources. Furthermore, low image resolution makes it difficult to accurately analyze detailed pathological features. This hinders AI models from making accurate diagnoses.

[0004] Therefore, with the current state of technology, a method capable of effectively processing the resolution and size of pathological images is needed. Resolution and size issues that arise during the analysis of pathological images by AI models are a major factor hindering diagnostic accuracy.

[0005] The purpose of the present disclosure is to provide a method for screening normal tissue using artificial intelligence, which determines with high accuracy whether a tissue sample is normal, and an electronic device for performing the same.

[0006] In one embodiment of the present disclosure, a method for screening normal tissue using artificial intelligence may be provided. The method may include the steps of: acquiring a full image corresponding to a tissue sample; acquiring a plurality of region-of-interest images based on the full image; acquiring a plurality of patches based on the plurality of region-of-interest images; extracting a plurality of signal values ​​using a first artificial intelligence model having the plurality of patches as inputs; acquiring a plurality of embedding vectors, each of which corresponds to each of the plurality of region-of-interest images, using a second artificial intelligence model having the plurality of signal values ​​and position values ​​corresponding to the plurality of patches as inputs; determining whether a region-of-interest image having a lesion exists among the plurality of region-of-interest images using a third artificial intelligence model having the plurality of embedding vectors as inputs; and determining that the tissue sample is normal based on determining that no region-of-interest image having a lesion exists among the plurality of region-of-interest images.

[0007] In one embodiment of the present disclosure, the first artificial intelligence model may be a model that has been pre-trained to output an image in which the stroma and epithelium portions of the normal tissue are distinguished by using images corresponding to normal tissue as input data.

[0008] In one embodiment of the present disclosure, the first artificial intelligence model may include at least one input layer that takes each of the plurality of patches as input and outputs a first feature vector, at least one downsampling layer that takes the first feature vector as input and outputs a second feature vector, at least one bottleneck layer that takes the second feature vector as input and outputs a third feature vector, at least one upsampling layer that takes the second feature vector and the third feature vector as inputs and outputs a fourth feature vector, and at least one output layer that takes the first feature vector and the fourth feature vector as inputs and outputs an image in which the stromal portion and the epithelial portion are distinguished within each of the plurality of patches.

[0009] In one embodiment of the present disclosure, each of the plurality of signal values ​​can be obtained from at least one bottleneck layer.

[0010] In one embodiment of the present disclosure, the second artificial intelligence model may be pre-trained to output an embedding vector including spatial relationships between the plurality of patches by performing attention between the plurality of patches.

[0011] In one embodiment of the present disclosure, the step of determining whether an image of a region of interest having a lesion exists among the plurality of region of interest images using a third artificial intelligence model having the plurality of embedding vectors as inputs may include the steps of: obtaining a lesion probability for each of the plurality of region of interest images using the third artificial intelligence model having the plurality of embedding vectors as inputs; determining whether the lesion probability exceeds a pre-defined first threshold value; identifying an image of a region of interest having the lesion probability exceeding the pre-defined first threshold value as an image of a region of interest having a lesion based on the determination that the lesion probability exceeds the pre-defined first threshold value; and determining whether an image of a region of interest having the lesion exists among the plurality of region of interest images based on the identification result.

[0012] In one embodiment of the present disclosure, the step of determining whether there is a region of interest image with a lesion among the plurality of region of interest images using a third artificial intelligence model having the plurality of embedding vectors as input may include the step of determining whether the lesion probability exceeds a second threshold value based on the determination that the lesion probability does not exceed the first threshold value, the step of identifying a region of interest image in which the lesion probability exceeds the second threshold value based on the determination that the lesion probability exceeds the second threshold value as a region of interest image in which the lesion probability exceeds the second threshold value based on the determination that the lesion probability does not exceed the second threshold value as a region of interest image without a lesion.

[0013] In one embodiment of the present disclosure, the plurality of signal values ​​may include information about the substrate portion and information about the epithelial portion of each of the plurality of patches.

[0014] In one embodiment of the present disclosure, the method may include a step of determining that the tissue sample is abnormal based on determining that an image of a region of interest having a lesion is present among the plurality of region of interest images.

[0015] In one embodiment of the present disclosure, an electronic device may be provided. The electronic device may include at least one processor including a processing circuit, and at least one memory storing at least one instruction. By causing the at least one processor to execute the at least one instruction, the electronic device may obtain a full image corresponding to a tissue sample, obtain a plurality of region-of-interest images based on the full image, obtain a plurality of patches based on the plurality of region-of-interest images, extract a plurality of signal values ​​using a first artificial intelligence model having the plurality of patches as inputs, obtain a plurality of embedding vectors each corresponding to each of the plurality of region-of-interest images using a second artificial intelligence model having the plurality of signal values ​​and position values ​​corresponding to the plurality of patches as inputs, determine whether a region-of-interest image with a lesion exists among the plurality of region-of-interest images using a third artificial intelligence model having the plurality of embedding vectors as inputs, and determine that the tissue sample is normal based on a determination that no region-of-interest image with a lesion exists among the plurality of region-of-interest images.

[0016] According to one embodiment of the present disclosure, by acquiring a full image corresponding to a tissue sample, acquiring a plurality of region-of-interest images, acquiring a plurality of patches, extracting a plurality of signal values ​​using a first artificial intelligence model, acquiring a plurality of embedding vectors using a second artificial intelligence model, and determining whether an image of a region-of-interest with a lesion exists using a third artificial intelligence model, the accuracy of artificial intelligence analysis due to problems with the resolution and size of pathological images can be improved. According to one embodiment of the present disclosure, by determining that a tissue sample is normal based on determining that an image of a region-of-interest with a lesion does not exist among the plurality of region-of-interest images, the reliability of pathological image analysis can be improved.

[0017] FIG. 1 is a conceptual diagram showing the operation of an electronic device according to one embodiment of the present disclosure.

[0018] FIG. 2 is a conceptual diagram showing the operation of a first image preprocessing module according to one embodiment of the present disclosure.

[0019] FIG. 3 is a conceptual diagram showing the operation of a second image preprocessing module according to one embodiment of the present disclosure.

[0020] FIG. 4 is a conceptual diagram showing the operation of a first artificial intelligence model according to one embodiment of the present disclosure.

[0021] FIGS. 5A and 5B are conceptual diagrams showing the structure and operation of a first artificial intelligence model according to one embodiment of the present disclosure.

[0022] FIG. 6 is a conceptual diagram showing the operation of a second artificial intelligence model according to one embodiment of the present disclosure.

[0023] FIG. 7 is a conceptual diagram showing the operation of a third artificial intelligence model according to one embodiment of the present disclosure.

[0024] FIG. 8 is a block diagram showing an electronic device according to one embodiment of the present disclosure.

[0025] FIG. 9 is a flowchart showing a method for screening normal tissue using artificial intelligence according to one embodiment of the present disclosure.

[0026] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. The embodiments will be described in sufficient detail to enable those skilled in the art to easily practice the present disclosure. However, the scope of the rights is not limited or restricted by these embodiments. Similar components in each drawing are designated by the same or similar reference numerals, and redundant descriptions of the same or similar components are omitted.

[0027] The terms used in the following description have been selected as common and universal in the relevant technical fields. However, other terms may be used depending on technological developments and / or changes, customs, and the preferences of technicians. Therefore, the terms used in the following description should not be construed as limiting the technical concepts, but rather as exemplary terms used to describe the embodiments.

[0028] Additionally, in certain cases, the applicant may arbitrarily select terms, in which case their detailed meanings will be described in the relevant description. Therefore, the terms used in the following description should be understood not simply as names, but based on their inherent meaning and the overall context of the specification.

[0029] Singular expressions may include plural expressions unless the context clearly indicates otherwise. Terms used herein, including technical or scientific terms, have the same meaning as commonly understood by a person of ordinary skill in the art described herein. Furthermore, terms containing ordinal numbers, such as "first" or "second," used herein may be used to describe various components, but such components should not be limited by such terms. Such terms are used solely to distinguish one component from another.

[0030] When a part of the specification is said to "include" a component, unless otherwise specifically stated, this does not exclude other components but rather implies the inclusion of other components. Furthermore, terms such as "part" and "module" used in the specification refer to a unit that processes at least one function or operation, which may be implemented in hardware, software, or a combination of hardware and software.

[0031] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the attached drawings so that those skilled in the art can easily practice the present disclosure. However, the present disclosure may be implemented in various different forms and is not limited to the embodiments described herein. In addition, in the drawings, parts that are not related to the description are omitted in order to clearly describe the present disclosure, and similar parts are designated with similar reference numerals throughout the specification. In addition, the reference numerals used in each drawing are only for the purpose of describing each drawing, and different reference numerals used in different drawings do not indicate different elements. The present disclosure will be described in detail below with reference to the attached drawings.

[0032] In the present disclosure, functions related to 'artificial intelligence' are operated through a processor and memory. The processor may be composed of one or more processors. In this case, one or more processors may be a general-purpose processor such as a CPU, AP, or DSP (Digital Signal Processor), a graphics-only processor such as a GPU or VPU (Vision Processing Unit), or an artificial intelligence-only processor such as an NPU. One or more processors control the processing of input data according to predefined operation rules or artificial intelligence models stored in memory. Alternatively, if one or more processors are artificial intelligence-only processors, the artificial intelligence-only processor may be designed with a hardware structure specialized for processing a specific artificial intelligence model.

[0033] The predefined operation rules or artificial intelligence models are characterized by being created through learning. Here, being created through learning means that the basic artificial intelligence model is trained using a learning algorithm using a plurality of learning data, thereby creating a predefined operation rules or artificial intelligence model set to perform a desired characteristic (or purpose). This learning may be performed on the device itself on which the artificial intelligence according to the present disclosure is performed, or may be performed through a separate server and / or system. Examples of the learning algorithm include, but are not limited to, supervised learning, unsupervised learning, semi-supervised learning, or reinforcement learning.

[0034] In the present disclosure, the 'artificial intelligence model' may be composed of a plurality of neural network layers. Each of the plurality of neural network layers has a plurality of weight values, and performs neural network operations through operations between the operation results of the previous layer and the plurality of weights. The plurality of weights of the plurality of neural network layers may be optimized by the learning results of the deep neural network model. For example, the plurality of weights may be updated so that the loss value or cost value obtained from the deep neural network model is reduced or minimized during the learning process. For example, the deep neural network model may include, but is not limited to, a convolutional neural network (CNN), a deep neural network (DNN), a recurrent neural network (RNN), a restricted boltzmann machine (RBM), a deep belief network (DBN), a bidirectional recurrent deep neural network (BRDNN), or deep Q-networks.

[0035] "Screening" may refer to the process of selecting or examining a specific subject. In one embodiment of the present disclosure, screening may refer to the process of using artificial intelligence to determine whether a tissue sample is normal.

[0036] A "tissue sample" may refer to a portion of tissue collected from a living organism. In one embodiment of the present disclosure, a tissue sample may refer to a portion of tissue used as input data for determining normality using an artificial intelligence model.

[0037] A "full image" may refer to an image that captures the entire appearance of a specific object. In one embodiment of the present disclosure, a full image may refer to an entire image corresponding to a tissue sample.

[0038] A "region of interest image" may refer to an image that highlights a specific portion of an entire image. In one embodiment of the present disclosure, a region of interest image may refer to an image of a specific region selected from an entire image of a tissue sample to determine whether there is a lesion.

[0039] A 'patch' may refer to a small piece of an image cut out from a specific portion of an image. In one embodiment of the present disclosure, a patch may refer to a small piece of an image extracted from an image of a region of interest.

[0040] The term "first artificial intelligence model" may refer to an encoder-decoder-based neural network model that generates a structural output similar to the input. In one embodiment of the present disclosure, the first artificial intelligence model may refer to a model that takes a tissue sample image as input and outputs an image divided into a stromal region and an epithelial region. The first artificial intelligence model may be pre-trained to infer a probability for the stromal region and a probability for the epithelial region for each pixel of the tissue sample image.

[0041] The "second AI model" may refer to an encoder model. In one embodiment of the present disclosure, the second AI model may refer to an encoder model with a transformer structure. The second AI model may output an embedding vector based on spatial correlations and patterns between multiple patches.

[0042] A "third artificial intelligence model" may refer to a model that predicts a label for a data group based on relationships between instances within the data group (also referred to as a batch or bag) and global group characteristics. In one embodiment of the present disclosure, the third artificial intelligence model may refer to a model that receives multiple images as input as a single group, selects images with lesions from among these images, or outputs a lesion probability.

[0043] 'Embedding vector' or 'feature vector' can refer to a vector that expresses data in numerical form.

[0044] A 'lesion' can mean an abnormal feature that occurs in a tissue or organ.

[0045] 'Substrate' can refer to the structure that supports the tissues or cells of an organism.

[0046] 'Epithelium' can refer to the layer of cells covering the surface of a living organism.

[0047] A "downsampling layer" may refer to a layer of a neural network that processes input data by reducing its size. In one embodiment of the present disclosure, the downsampling layer may refer to a layer that processes feature vectors by reducing their size within the first artificial intelligence model.

[0048] The term "bottleneck layer" may refer to the deepest layer in a neural network. In one embodiment of the present disclosure, the bottleneck layer may refer to a layer that extracts signal values ​​by maximally compressing information about patches within the first artificial intelligence model.

[0049] An "upsampling layer" may refer to a layer of a neural network that enlarges and processes input data. In one embodiment of the present disclosure, the upsampling layer may refer to a layer that enlarges and processes a reduced feature vector within the first artificial intelligence model.

[0050] "Attention" may refer to a technique for processing input data by focusing on a specific portion of the data. In one embodiment of the present disclosure, attention may refer to a technique used by a second AI model to generate an embedding vector by considering the spatial relationships between patches.

[0051] FIG. 1 is a conceptual diagram showing the operation of an electronic device according to one embodiment of the present disclosure.

[0052] Referring to FIG. 1, an electronic device (1000) can preprocess an entire image corresponding to a tissue sample and utilize multiple artificial intelligence models to determine whether the tissue sample is normal tissue. The electronic device (1000) may include an image preprocessing module (1100), an artificial intelligence module (1200), and a normal tissue determination module (1300). For example, the electronic device (1000) may be a smartphone, a tablet, a laptop, a desktop computer, a server, an embedded system, an industrial computer, or a medical device. However, the present disclosure is not limited thereto.

[0053] The electronic device (1000) can preprocess an entire image corresponding to a tissue sample through an image preprocessing module (1100). The electronic device (1000) can provide the preprocessed image to an artificial intelligence module (1200). The electronic device (1000) can extract region-of-interest images based on the entire image. The electronic device (1000) can segment each region-of-interest image into a plurality of patches.

[0054] The electronic device (1000) can analyze a preprocessed image using multiple artificial intelligence models via the artificial intelligence module (1200). The electronic device (1000) can transmit the analysis results to the normal tissue determination module (1300). Based on the analysis results received from the artificial intelligence module (1200) via the normal tissue determination module (1300), the electronic device (1000) can determine whether the tissue sample is normal tissue. The electronic device (1000) can output the final determination result.

[0055] The image preprocessing module (1100) can preprocess the entire image corresponding to the tissue sample. The image preprocessing module (1100) can extract region-of-interest images based on the entire image and segment each region-of-interest image into multiple patches. In one embodiment of the present disclosure, the image preprocessing module (1100) can perform tasks such as filtering, noise removal, contrast adjustment, resizing, rotation, color correction, and transformation. However, the present disclosure is not limited thereto.

[0056] In one embodiment of the present disclosure, the image preprocessing module (1100) can selectively apply various preprocessing algorithms. For example, an appropriate preprocessing method can be selected and applied based on the characteristics of a specific tissue sample.

[0057] The artificial intelligence module (1200) can perform a function of analyzing an image of a tissue sample within the electronic device (1000). The artificial intelligence module (1200) can include multiple artificial intelligence models. For example, the artificial intelligence module (1200) can be a convolutional neural network (CNN), a recurrent neural network (RNN), a deep neural network (DNN), a generative adversarial network (GAN), a reinforcement learning model, a support vector machine (SVM), or a random forest. However, the present disclosure is not limited thereto.

[0058] The artificial intelligence module (1200) can receive a preprocessed image (patch) from the image preprocessing module (1100). The artificial intelligence module (1200) can analyze the preprocessed image and provide the analysis result to the normal tissue judgment module (1300). The artificial intelligence module (1200) can extract the signal value of the preprocessed image by utilizing multiple artificial intelligence models. The artificial intelligence module (1200) can obtain an embedding vector by inputting the signal value and the location value corresponding to the patch. The artificial intelligence module (1200) can determine whether an image of a region of interest containing a lesion exists using the embedding vector.

[0059] The normal tissue judgment module (1300) can determine whether a tissue sample is normal tissue based on the analysis result received from the artificial intelligence module (1200). The normal tissue judgment module (1300) can output the final judgment result of the electronic device (1000). The normal tissue judgment module (1300) can determine the tissue sample as at least one of normal, abnormal, and unknown based on the analysis result. For example, the normal tissue judgment module (1300) can make a judgment based on predefined conditions. For example, the normal tissue judgment module (1300) can make a judgment using an artificial intelligence model. For example, the normal tissue judgment module (1300) can use techniques related to a neural network, a support vector machine (SVM), a random forest, a decision tree, k-nearest neighbors (k-NN), logistic regression, and naive Bayes. However, the present disclosure is not limited thereto.

[0060] The normal tissue judgment module (1300) can receive as input a plurality of embedding vectors received from the artificial intelligence module (1200) and determine whether an image of a region of interest containing a lesion exists. If the normal tissue judgment module (1300) determines that an image of a region of interest containing a lesion does not exist, the normal tissue judgment module (1300) can determine that the tissue sample is normal. The normal tissue judgment module (1300) can output this judgment result as the final judgment result of the electronic device (1000).

[0061] FIG. 2 is a conceptual diagram illustrating the operation of a first image preprocessing module according to one embodiment of the present disclosure. Regarding the image preprocessing module (1100), any details that overlap with those described in FIG. 1 will be omitted.

[0062] Referring to FIG. 2 together with FIG. 1, the image preprocessing module (1100) can preprocess the entire image corresponding to the tissue sample. The image preprocessing module (1100) can include a first image preprocessing module (1110). The first image preprocessing module (1110) can receive the entire image as input and extract a plurality of region-of-interest images. In one embodiment of the present disclosure, the first image preprocessing module (1110) can extract region-of-interest images from the entire image using an artificial intelligence model or based on user input.

[0063] The first image preprocessing module (1110) can receive an entire image corresponding to a tissue sample and extract multiple region-of-interest images. In one embodiment of the present disclosure, the first image preprocessing module (1110) can extract multiple region-of-interest images from the entire image according to predefined conditions. In one embodiment of the present disclosure, the first image preprocessing module (1110) can identify and extract region-of-interest images from the entire image by utilizing an artificial intelligence model. For example, the first image preprocessing module (1110) can identify a region with a specific pattern or characteristic as a region-of-interest from the entire image of the tissue sample by utilizing the artificial intelligence model. However, the present disclosure is not limited thereto. In one embodiment of the present disclosure, the first image preprocessing module (1110) can extract multiple region-of-interest images from the entire image based on a user input corresponding to region selection.

[0064] The first image preprocessing module (1110) can preprocess the entire image corresponding to the tissue sample to extract multiple region-of-interest images. Through this process, the first image preprocessing module (1110) can provide a basis for determining the normality of the tissue sample with high accuracy.

[0065] In one embodiment of the present disclosure, the first image preprocessing module (1110) can selectively apply various artificial intelligence models to extract region of interest images. The first image preprocessing module (1110) can extract a region of interest suitable for a specific type of tissue sample using a specific artificial intelligence model. The first image preprocessing module (1110) can extract a region of interest suitable for a different type of tissue sample using a different artificial intelligence model. In this manner, the first image preprocessing module (1110) can flexibly respond to various tissue samples.

[0066] FIG. 3 is a conceptual diagram illustrating the operation of a second image preprocessing module according to one embodiment of the present disclosure. Regarding the image preprocessing module (1100), any details that overlap with those described in FIGS. 1 and 2 will be omitted.

[0067] Referring to FIG. 3, along with FIGS. 1 and 2, the image preprocessing module (1100) may include a second image preprocessing module (1120). The second image preprocessing module (1120) may receive region of interest images from the first image preprocessing module (1110) and divide them into a plurality of patches.

[0068] The second image preprocessing module (1120) may perform a function of dividing region-of-interest images into multiple patches. For example, the size of the region-of-interest image may be 7680x7860, and each patch may be 480x480. However, the present disclosure is not limited thereto.

[0069] The second image preprocessing module (1120) can segment each region of interest image into multiple patches. These segmented patches can then be used as input for an artificial intelligence model.

[0070] In one embodiment of the present disclosure, the second image preprocessing module (1120) can segment the region of interest images into patches of various sizes. For example, the patch sizes may be 480x480, 240x240, 120x120, 60x60, 30x30, 15x15, or 7x7. However, the present disclosure is not limited thereto. In one embodiment of the present disclosure, the patch sizes may vary depending on the settings of the user or manufacturer, and a predefined size may be preset for each tissue. These patches of various sizes can contribute to improving the learning and prediction accuracy of the artificial intelligence model, and can more precisely determine whether the tissue sample is normal.

[0071] Figure 4 is a conceptual diagram illustrating the operation of a first artificial intelligence model according to one embodiment of the present disclosure. Regarding the artificial intelligence module (1200), any details that overlap with those described in Figure 1 will be omitted.

[0072] Referring to FIG. 4 together with FIG. 1, the artificial intelligence module (1200) can preprocess an entire image corresponding to a tissue sample and determine whether the tissue sample is normal tissue by utilizing multiple artificial intelligence models. The artificial intelligence module (1200) can include a first artificial intelligence model (1210). The first artificial intelligence model (1210) can receive multiple patches (PATCH) corresponding to each of the region of interest images from the image preprocessing module (1100). The electronic device (1000) can extract a signal value (SIG) corresponding to each patch by utilizing the first artificial intelligence model (1210) that receives multiple patches (PATCH) as input. The signal value (SIG) can be classified into a normal signal and an abnormal signal according to predefined conditions. The abnormal signals can be grouped to form an abnormal region (AR) within the region of interest image.

[0073] In one embodiment of the present disclosure, the first artificial intelligence model (1210) may be an image segmentation model. For example, the first artificial intelligence model (1210) may be a model pre-trained to output an image in which the stroma and the epithelium are distinguished by using images corresponding to normal tissue as input data. In one embodiment of the present disclosure, the artificial intelligence module (1200) may obtain a plurality of signal values ​​(SIG) by utilizing only the encoder part of the first artificial intelligence model (1210). The artificial intelligence module (1200) may obtain the signal value (SIG) based on a feature vector which is an output of the deepest layer of the first artificial intelligence model (1210).

[0074] The first artificial intelligence model (1210) can receive as input multiple patches corresponding to the region of interest image received from the image preprocessing module (1100), and extract a signal value (SIG) corresponding to each patch. The signal value (SIG) can be classified into a normal signal and an abnormal signal according to predefined conditions.

[0075] Figures 5a and 5b are conceptual diagrams illustrating the structure and operation of a first artificial intelligence model according to one embodiment of the present disclosure. Regarding the first artificial intelligence model (1210), any details that overlap with those described in Figure 4 will be omitted.

[0076] Referring to FIGS. 5A and 5B along with FIGS. 1 and 4, the first artificial intelligence model (1210) may be a model designed to determine whether a tissue sample is normal. The first artificial intelligence model (1210) may be composed of at least one input layer (IL), at least one downsampling layer (DL), at least one bottleneck layer (BL), at least one upsampling layer (UL), and at least one output layer (OL). The input layer (IL) may receive a patch of a tissue sample as input and generate a first feature vector. The downsampling layer (DL) may reduce the first feature vector to generate a second feature vector. The bottleneck layer (BL) may extract a third feature vector based on the second feature vector. The upsampling layer (UL) may expand the third feature vector to generate a fourth feature vector. The output layer (OL) may receive the first feature vector and the fourth feature vector as input and ultimately generate a segmentation map for the patch.

[0077] The input layer (IL) can receive a patch of tissue sample and output a first feature vector. The input layer (IL) can extract features of the input patch through a convolution operation. The input layer (IL) is the initial layer of the first artificial intelligence model (1210) and can pass the first feature vector to the downsampling layer (DL).

[0078] In one embodiment of the present disclosure, the input layer (IL) may have a structure capable of performing various types of convolution operations. For example, the input layer (IL) may perform 2D convolution, 3D convolution, deep convolution, or multi-channel convolution. However, the present disclosure is not limited thereto. The input layer (IL) may have a flexible structure capable of processing various types of patch inputs.

[0079] A downsampling layer (DL) can reduce the input first feature vector and output a second feature vector. For example, the downsampling layer (DL) can use methods such as max pooling, average pooling, and strided convolution. However, the present disclosure is not limited thereto.

[0080] The bottleneck layer (BL) is the deepest layer in the first artificial intelligence model (1210) and can extract key features of input data. The bottleneck layer (BL) can operate between the downsampling layer (DL) and the upsampling layer (UL). The bottleneck layer (BL) can receive a second feature vector as input and output a third feature vector. For example, the bottleneck layer (BL) can extract important information by reducing the dimensionality of the input feature vector through a convolution operation. However, the present disclosure is not limited thereto.

[0081] The third feature vector may correspond to a signal value. Since the first artificial intelligence model (1210) is trained to output probability values ​​of the epithelial portion and the stromal portion (or additionally the probability value of the background portion) from normal tissue patches, when an abnormal tissue patch is input to the first artificial intelligence model (1210), it may output a signal (SIG_A) having a different value from a signal (SIG_N) having a normal signal value. When the tissue patch is normal, the bottleneck layer (BL) may output a signal (SIG_N) having a normal signal value. When the tissue patch is abnormal, the bottleneck layer (BL) may output a signal (SIG_A) having an abnormal signal value.

[0082] The upsampling layer (UL) can receive the second feature vector and the third feature vector as inputs and output a fourth feature vector. The upsampling layer (UL) can concatenate the second feature vector with the third feature vector. The upsampling layer (UL) can expand the third feature vector and output a fourth feature vector. The upsampling layer (UL) can perform upconvolution on the third feature vector. The fourth feature vector can be passed to the output layer (OL).

[0083] The output layer (OL) can ultimately generate a probability map for the patch. The output layer (OL) can receive the first feature vector and the fourth feature vector as input and output probability values ​​for the stromal portion and the epithelial portion within the plurality of patches. For example, if the probability value for the stromal portion in the first pixel is higher than the probability value for the epithelial portion, the first pixel can be estimated as the stromal portion. For example, if the probability value for the stromal portion in the second pixel is lower than or equal to the probability value for the epithelial portion, the second pixel can be estimated as the epithelial portion. The upsampling layer (UL) can concatenate the first feature vector with the fourth feature vector. For example, the output layer (OL) can process the input feature vectors through a convolution operation to generate a final segmentation map. However, the present disclosure is not limited thereto.

[0084] In one embodiment of the present disclosure, the output layer (OL) can output an image having RGB channels. According to one embodiment of the present disclosure, the output layer (OL) can output an image in which a stromal portion and an epithelial portion are visually distinguished within a normal tissue. For example, a portion estimated as a stromal portion within a patch may have a value of at least one first channel (e.g., an R channel) among the RGB channels set to a first value (e.g., 255). For example, a portion estimated as an epithelial portion within a patch may have a value of at least one second channel (e.g., a G channel and a B channel) excluding at least one first channel among the RGB channels set to a second value. However, the present disclosure is not limited thereto, and at least one value within the RGB channels may be arbitrarily adjusted so that the output layer (OL) outputs an image in which a stromal portion and an epithelial portion are visually distinguished.

[0085] In one embodiment of the present disclosure, the output layer (OL) may have a flexible structure that can be applied to various artificial intelligence models. The output layer (OL) may be replaced with the output layer of another artificial intelligence model, thereby enabling analysis of various tissue samples. This structural flexibility allows the output layer (OL) to be applied to a variety of applications.

[0086] Figure 6 is a conceptual diagram illustrating the operation of a second artificial intelligence model according to one embodiment of the present disclosure. Regarding the artificial intelligence module (1200), any details that overlap with those described in Figures 1 and 4 will be omitted.

[0087] Referring to FIG. 6, along with FIGS. 1 and 4, the artificial intelligence module (1200) may include a second artificial intelligence model (1220). The second artificial intelligence model (1220) may receive as input a plurality of signal values ​​and a plurality of position values ​​corresponding to a plurality of patches of a specific region of interest image, and output an embedding vector corresponding to the corresponding region of interest image.

[0088] In one embodiment of the present disclosure, the artificial intelligence module (1200) can extract multiple signal values ​​from multiple patches using a first artificial intelligence model. One patch can correspond to one signal value and one location value. The location value can refer to a coordinate value of the corresponding patch on a region of interest image. The second artificial intelligence model (1220) can obtain multiple embedding vectors, each corresponding to a respective one of the multiple region of interest images, by inputting multiple signal values ​​and location values ​​corresponding to the multiple patches.

[0089] The second artificial intelligence model (1220) can receive signal values ​​and position values ​​corresponding to multiple patches as input and output an embedding vector. For example, the second artificial intelligence model (1220) may be a transformer, a recurrent neural network (RNN), a convolutional neural network (CNN), a long short-term memory (LSTM), a gated recurrent unit (GRU), a graph neural network (GNN), or a deep neural network (DNN). However, the present disclosure is not limited thereto.

[0090] For example, the second artificial intelligence model (1220) may be an encoder model having a transformer structure. The second artificial intelligence model (1220) may output an embedding vector that includes spatial relationships between multiple patches by performing attention between multiple patches.

[0091] Figure 7 is a conceptual diagram illustrating the operation of a third artificial intelligence model according to one embodiment of the present disclosure. Regarding the artificial intelligence module (1200), any details that overlap with those described in Figures 1, 4, and 6 will be omitted.

[0092] Referring to FIG. 7, along with FIGS. 1, 4, and 6, the artificial intelligence module (1200) may include a third artificial intelligence model (1230). The third artificial intelligence model (1230) may receive embedding vectors corresponding to each region of interest image, and output a lesion probability for each region of interest image.

[0093] The third artificial intelligence model (1230) is a subcomponent of the artificial intelligence module (1200) and can be used to determine the normality of a tissue sample. For example, the third artificial intelligence model (1230) may be a convolutional neural network (CNN), a recurrent neural network (RNN), a deep neural network (DNN), a generative adversarial network (GAN), a support vector machine (SVM), a K-nearest neighbors (KNN), or a decision tree. However, the present disclosure is not limited thereto.

[0094] In one embodiment of the present disclosure, the third artificial intelligence model (1230) may be a model trained using the Multiple Instance Learning (MIL) method. The MIL method is an approach that divides an image of a tissue sample into several small patches, generates an embedding vector for each patch, and calculates a lesion probability based on these embedding vectors. The third artificial intelligence model (1230) estimates the lesion probability for each region of interest image in this manner, and can ultimately determine whether the tissue sample is normal or abnormal. This MIL-based learning can learn the contribution of individual patches using only the label information of the entire image, thereby improving the accuracy and reliability of lesion detection. The third artificial intelligence model (1230) can effectively learn the characteristics of lesions and can contribute to increasing clinical applicability.

[0095] The third artificial intelligence model (1230) can receive multiple embedding vectors as input and calculate a lesion probability for each region of interest image. The electronic device (1000) can determine whether the lesion probability exceeds a predefined threshold. If the lesion probability exceeds the threshold, the electronic device (1000) can identify the region of interest image as an image containing a lesion. If the lesion probability does not exceed the threshold, the electronic device (1000) can identify the image as an image without a lesion. Through this process, the normality of a tissue sample can be determined with high accuracy.

[0096] In one embodiment of the present disclosure, the electronic device (1000) may receive a plurality of embedding vectors as input and determine whether the lesion probability exceeds a predefined first threshold. If the lesion probability exceeds the first threshold, the electronic device (1000) may identify the corresponding image as an image with a lesion. If the lesion probability does not exceed the first threshold, the electronic device (1000) may additionally determine whether the lesion probability exceeds a predefined second threshold. If the lesion probability exceeds the second threshold, the electronic device (1000) may identify the corresponding image as an image for which the presence or absence of a lesion cannot be determined. If the lesion probability does not exceed the second threshold, the electronic device (1000) may identify the corresponding image as an image without a lesion. In this manner, the third artificial intelligence model (1230) may more accurately determine whether a tissue sample is normal.

[0097] FIG. 8 is a block diagram showing an electronic device according to an embodiment of the present disclosure. With respect to the image preprocessing module (1100), the artificial intelligence module (1200), the first artificial intelligence model (1210), the second artificial intelligence model (1220), the third artificial intelligence model (1230), and the normal tissue determination module (1300), any overlapping content with that described in FIGS. 1, 2, 3, 4, 6, 7, and 5A will be omitted. The configuration, function, and operation of the electronic device (2000) may correspond to the configuration, function, and operation of the electronic device (1000) of FIG. 1.

[0098] Referring to FIG. 8 along with FIG. 1, the electronic device (2000) may include a processor (2100), storage (2200), a communication interface (2300), and memory (2400). For example, the electronic device (2000) may be a smartphone, a tablet, a computer, a server, network equipment, a medical device, or industrial equipment. However, the present disclosure is not limited thereto.

[0099] The electronic device (2000) can preprocess the entire image corresponding to the tissue sample and utilize multiple artificial intelligence models to determine whether the tissue sample is normal. The electronic device (2000) can determine whether the tissue sample is normal with high accuracy.

[0100] The processor (2100) may be a core component capable of performing various operations within the electronic device (2000). The processor (2100) may play a key role in cooperating with other components within the electronic device (2000) to process and analyze images of tissue samples.

[0101] The processor (2100) can store and execute at least one instruction in cooperation with the memory (2400). The processor (2100) can preprocess an image of a tissue sample through the image preprocessing module (1100) and the artificial intelligence module (1200), and utilize multiple artificial intelligence models to determine whether the tissue sample is normal tissue. In this process, the processor (2100) can extract multiple signal values ​​and obtain multiple embedding vectors.

[0102] In one embodiment of the present disclosure, the processor (2100) can collaborate with a first artificial intelligence model (1210), a second artificial intelligence model (1220), and a third artificial intelligence model (1230) to determine with high accuracy whether a tissue sample is normal. The processor (2100) can perform attention between a plurality of patches through the second artificial intelligence model (1220) that is pre-trained to output an embedding vector including a spatial relationship between the plurality of patches. The processor (2100) can determine whether an image of a region of interest containing a lesion exists using the third artificial intelligence model (1230) that inputs the plurality of embedding vectors.

[0103] The processor (2100) can play a key role in analyzing images of tissue samples and determining the presence of lesions by collaborating with various artificial intelligence models within the electronic device (2000). The processor (2100) can obtain multiple embedding vectors by inputting multiple signal values ​​and position values ​​corresponding to multiple patches. The processor (2100) can obtain lesion probabilities by inputting multiple embedding vectors and determine whether the lesion probabilities exceed a predefined threshold.

[0104] Storage (2200) may be a device capable of storing and managing data. For example, storage (2200) may be a hard disk drive (HDD), a solid state drive (SSD), flash memory, an optical disk drive, network storage, cloud storage, or a USB drive. However, the present disclosure is not limited thereto.

[0105] Storage (2200) can store data necessary for preprocessing an entire image corresponding to a tissue sample within the electronic device (2000) and for determining whether the tissue sample is normal tissue using multiple artificial intelligence models. Storage (2200) can store instructions to be executed by the processor (2100). Storage (2200) can work with memory (2400) to store training data and result data of the artificial intelligence model.

[0106] In one embodiment of the present disclosure, the storage (2200) may store a plurality of region-of-interest images and a plurality of patches, thereby providing data necessary for the artificial intelligence module (1200) to extract a plurality of signal values ​​and embedding vectors. The storage (2200) may store data necessary for the training and execution of the first artificial intelligence model (1210), the second artificial intelligence model (1220), and the third artificial intelligence model (1230). The storage (2200) may store data necessary for the normal tissue judgment module (1300) to determine with high accuracy whether a tissue sample is normal.

[0107] The communication interface (2300) may be a component that enables data transmission and reception between the electronic device (2000) and an external device. For example, the communication interface (2300) may be wired communication, wireless communication, short-range communication, long-range communication, Internet Protocol (IP) communication, Bluetooth communication, or near field communication (NFC). However, the present disclosure is not limited thereto.

[0108] The communication interface (2300) can support data transmission between the processor (2100) and the memory (2400) within the electronic device (2000). The communication interface (2300) enables the electronic device (2000) to be connected to an external network and transmit and receive data. The communication interface (2300) enables the electronic device (2000) to receive or transmit images of a tissue sample through communication with an external device.

[0109] In one embodiment of the present disclosure, the communication interface (2300) can receive additional learning data through a connection to an external database while the electronic device (2000) utilizes multiple artificial intelligence models to determine whether a tissue sample is normal. The communication interface (2300) enables the electronic device (2000) to update the artificial intelligence model through communication with an external server. The communication interface (2300) can transmit the analysis results of the tissue sample to the external device based on the data received by the electronic device (2000) from the external device.

[0110] Memory (2400) is a component capable of storing and managing data within an electronic device (2000). Memory (2400) may be implemented as various types of memory devices. For example, memory (2400) may be random access memory (RAM), read-only memory (ROM), flash memory, dynamic random access memory (DRAM), static random access memory (SRAM), non-volatile memory (NVM), and cache memory. However, the present disclosure is not limited thereto.

[0111] The memory (2400) can support the operation of various modules, such as the normal tissue judgment module (1300) and the artificial intelligence module (1200) of the electronic device (2000). The memory (2400) can efficiently store and retrieve data to improve the overall function of the electronic device (2000).

[0112] In one embodiment of the present disclosure, the memory (2400) can provide additional storage capacity to expand the functionality of the electronic device (2000). The memory (2400) can store various types of data and support high-speed data access to optimize the performance of the electronic device (2000). The memory (2400) can ensure data integrity while maintaining compatibility with various components of the electronic device (2000).

[0113] The image preprocessing module (1100) can preprocess the entire image corresponding to the tissue sample. The image preprocessing module (1100) can contribute to determining whether the tissue sample is normal tissue by utilizing multiple artificial intelligence models within the electronic device (2000). The electronic device (2000) includes a processor (2100) and a memory (2400), and can execute and store at least one instruction.

[0114] In one embodiment of the present disclosure, the image preprocessing module (1100) can acquire multiple region-of-interest images based on an entire image. The image preprocessing module (1100) can segment each of the multiple region-of-interest images into multiple patches.

[0115] The artificial intelligence module (1200) can preprocess the entire image corresponding to a tissue sample and utilize multiple artificial intelligence models to determine whether the tissue sample is normal tissue. The artificial intelligence module (1200) may include a first artificial intelligence model (1210), a second artificial intelligence model (1220), and a third artificial intelligence model (1230).

[0116] In one embodiment of the present disclosure, the first AI model (1110) can generate signal values ​​including information about the stromal and epithelial portions of each of the plurality of patches. The second AI model (1220) can output an embedding vector including spatial relationships between the patches. The third AI model (1230) can identify an image of a region of interest containing a lesion by determining whether the lesion probability exceeds a predefined threshold. This process can contribute to determining the normality of a tissue sample with high accuracy.

[0117] The first artificial intelligence model (1210) can receive multiple patches as input and extract multiple signal values. The first artificial intelligence model (1210) may be a model that has been pre-trained to output an image in which a stroma portion and an epithelium portion are distinguished by using images corresponding to normal tissue as input data. The first artificial intelligence model (1210) may output a first feature vector through at least one input layer. The first artificial intelligence model (1210) may include at least one downsampling layer that takes the first feature vector as input and outputs a second feature vector. The first artificial intelligence model (1210) may include at least one bottleneck layer that takes the second feature vector as input and outputs a third feature vector. The first artificial intelligence model (1210) may include at least one upsampling layer that takes the second feature vector and the third feature vector as input and outputs a fourth feature vector. The first artificial intelligence model (1210) may include at least one output layer that takes the first feature vector and the fourth feature vector as inputs and outputs an image in which the stromal portion and the epithelial portion are distinguished based on probability values ​​and / or probability values ​​for the stromal portion and the epithelial portion within a plurality of patches.

[0118] In one embodiment of the present disclosure, the first artificial intelligence model (1210) may generate a plurality of signal values ​​including information about the substrate portion and the epithelial portion of each of the plurality of patches. Each of the plurality of signal values ​​may correspond to a third feature vector output by at least one bottleneck layer.

[0119] The second artificial intelligence model (1220) can perform attention between multiple patches and output an embedding vector including the spatial relationship between the multiple patches. The second artificial intelligence model (1220) can receive multiple signal values ​​and position values ​​corresponding to the multiple patches as input and obtain multiple embedding vectors, each corresponding to a respective one of the multiple region-of-interest images. The second artificial intelligence model (1220) can be trained in advance to output an embedding vector including the spatial relationship between the multiple patches.

[0120] In one embodiment of the present disclosure, the third artificial intelligence model (1230) can receive a plurality of embedding vectors as input and output a lesion probability for each of a plurality of region-of-interest images.

[0121] The normal tissue judgment module (1300) can determine whether a tissue sample in the entire image is normal tissue. The normal tissue judgment module (1300) can receive a lesion probability from the third artificial intelligence model (1230). The normal tissue judgment module (1300) can determine whether the lesion probability exceeds a predefined threshold value and identify the corresponding region of interest image as a region of interest image with a lesion. For example, if the lesion probability does not exceed a predefined first threshold value, the normal tissue judgment module (1300) can additionally determine whether the lesion probability exceeds a predefined second threshold value and identify the corresponding region of interest image as a region of interest image in which the presence or absence of a lesion cannot be confirmed. If the lesion probability does not exceed the second threshold value, the normal tissue judgment module (1300) can identify the corresponding region of interest image as a region of interest image without a lesion.

[0122] In one embodiment of the present disclosure, the normal tissue judgment module (1300) may determine whether the lesion probability exceeds a predefined threshold. If the lesion probability exceeds the threshold, the normal tissue judgment module (1300) may identify the corresponding region of interest image as a region of interest image containing a lesion.

[0123] In one embodiment of the present disclosure, if the lesion probability does not exceed a threshold value, the normal tissue determination module (1300) can identify the corresponding region of interest image as a region of interest image without a lesion. This process can determine the normality of a tissue sample with high accuracy.

[0124] FIG. 9 is a flowchart showing a method for screening normal tissue using artificial intelligence according to one embodiment of the present disclosure.

[0125] Referring to FIG. 9 along with FIG. 1, at step S910, the electronic device (1000) can acquire a full image corresponding to a tissue sample. The full image may also be referred to as a whole slide image (WSI). The electronic device (1000) can capture a full image of the tissue sample from a high-resolution imaging sensor. For example, the electronic device (1000) can acquire an image from an optical microscope, an electron microscope, or a digital scanner. However, the present disclosure is not limited thereto.

[0126] In one embodiment of the present disclosure, step S910 may include acquiring images of a tissue sample from various angles. The electronic device (1000) can use multi-angle imaging technology to more accurately analyze the three-dimensional structure of the tissue sample.

[0127] At step S920, the electronic device (1000) can acquire multiple region-of-interest images based on the entire image. The electronic device (1000) can identify a specific region of interest from the entire image using an image processing algorithm. For example, the electronic device (1000) can select a region of interest based on color, contrast, or texture. However, the present disclosure is not limited thereto.

[0128] In one embodiment of the present disclosure, step S920 may include adjusting the size and shape of a region of interest. The electronic device (1000) can select regions of interest of various sizes and shapes to perform more detailed analysis.

[0129] At step S930, the electronic device (1000) can acquire multiple patches based on multiple region-of-interest images. The electronic device (1000) can divide the region-of-interest image into small patches and individually analyze each patch. For example, the electronic device (1000) can generate patches using a grid pattern of a certain size. However, the present disclosure is not limited thereto.

[0130] In one embodiment of the present disclosure, step S930 may include adjusting the resolution of the patch. For example, the electronic device (1000) may use an upscaling model to change the resolution of the patch to a higher resolution. The electronic device (1000) may generate a high-resolution patch, thereby enabling more accurate analysis.

[0131] At step S940, the electronic device (1000) can extract a plurality of signal values ​​using a first artificial intelligence model that inputs a plurality of patches. The electronic device (1000) can generate a signal value having substrate and epithelial information for each patch through the first artificial intelligence model.

[0132] In one embodiment of the present disclosure, step S940 may include a signal value normalization step. The electronic device (1000) may normalize the signal value to increase the consistency of the analysis.

[0133] In step S950, the electronic device (1000) can obtain a plurality of embedding vectors corresponding to each of the plurality of region-of-interest images using a second artificial intelligence model that inputs a plurality of signal values ​​and position values ​​corresponding to a plurality of patches. The electronic device (1000) can generate an embedding vector using the second artificial intelligence model as input data consisting of signal values ​​and position values ​​on a two-dimensional coordinate system.

[0134] In one embodiment of the present disclosure, step S950 may include a step of reducing the dimensionality of the embedding vector. The electronic device (1000) may reduce the complexity of data using a dimensionality reduction technique.

[0135] At step S960, the electronic device (1000) can determine whether a region of interest image containing a lesion exists among the plurality of region of interest images using a third artificial intelligence model that inputs the plurality of embedding vectors. The electronic device (1000) can analyze the embedding vectors to determine whether a lesion exists. In one embodiment of the present disclosure, the electronic device (1000) can identify an image containing a lesion using a classification model or algorithm. However, the present disclosure is not limited thereto.

[0136] In one embodiment of the present disclosure, step S960 may include calculating a lesion probability. The electronic device (1000) can quantify the likelihood of a lesion's presence to provide a more precise diagnosis.

[0137] At step S970, the electronic device (1000) may determine that the tissue sample is normal based on the determination that no region of interest image with a lesion exists among the multiple region of interest images. The electronic device (1000) may assess the condition of the tissue sample based on the image without a lesion. For example, the electronic device (1000) may classify the tissue sample as normal if all region of interest images are determined to be normal. However, the present disclosure is not limited thereto.

[0138] In one embodiment of the present disclosure, step S970 may include reporting the status of a tissue sample. The electronic device (1000) may provide analysis results to the user to assist in the diagnostic process.

[0139] A method according to one embodiment of the present disclosure may be provided as a computer program product. The computer program product may be traded as a commodity between a seller and a buyer. The computer program product may be distributed in the form of a machine-readable storage medium (e.g., a compact disc read-only memory (CD-ROM)), or may be distributed online (e.g., downloaded or uploaded) through an application store or directly between two user devices (e.g., smartphones). In the case of online distribution, at least a portion of the computer program product (e.g., a downloadable app) may be temporarily stored or temporarily generated in a machine-readable storage medium, such as the memory of a manufacturer's server, an application store's server, or an intermediary server.

[0140] In the embodiments described above, components according to the technical concept of the present disclosure have been described using terms such as "first," "second," "third," etc. However, terms such as "first," "second," "third," etc. are used to distinguish components from each other and do not limit the technical concept of the present disclosure. Terms such as "first," "second," "third," etc. do not imply a numerical meaning of order or any form.

[0141] The embodiments described above are specific examples for implementing the present disclosure. It should be understood that the present disclosure encompasses not only the embodiments described above, but also embodiments that can be simply designed or easily modified using the embodiments described above. Therefore, the scope of the present disclosure should not be limited to the embodiments described above, but should be defined not only by the claims described below but also by equivalents of the claims of this invention.

Claims

1. A method for screening normal tissue using artificial intelligence, A step of acquiring a full image corresponding to a tissue sample; A step of acquiring multiple images of a region of interest based on the entire image; A step of obtaining a plurality of patches based on the plurality of region of interest images; A step of extracting a plurality of signal values ​​using a first artificial intelligence model that takes the plurality of patches as input; A step of obtaining a plurality of embedding vectors, each corresponding to each of the plurality of interest region images, using a second artificial intelligence model that inputs the plurality of signal values ​​and the position values ​​corresponding to the plurality of patches; and A step of determining whether an image of a region of interest with a lesion exists among the plurality of region of interest images using a third artificial intelligence model that takes the plurality of embedding vectors as input; A method comprising the step of determining that the tissue sample is normal based on determining that no region of interest image with a lesion exists among the plurality of region of interest images.

2. In paragraph 1, The above first artificial intelligence model is a model that has been pre-trained to output an image in which the stroma and epithelium portions of the normal tissue are distinguished by using images corresponding to normal tissue as input data.

3. In paragraph 2, The above first artificial intelligence model: At least one input layer that takes each of the plurality of patches as input and outputs a first feature vector; At least one downsampling layer that takes the first feature vector as input and outputs a second feature vector; At least one bottleneck layer that takes the second feature vector as input and outputs a third feature vector; At least one upsampling layer that takes the second feature vector and the third feature vector as inputs and outputs a fourth feature vector; and A method comprising at least one output layer that outputs an image in which the substrate portion and the epithelium portion are distinguished within each of the plurality of patches by using the first feature vector and the fourth feature vector as inputs.

4. In paragraph 3, A method wherein each of the plurality of signal values ​​is obtained from at least one bottleneck layer.

5. In paragraph 1, A method in which the second artificial intelligence model is pre-trained to output an embedding vector including spatial relationships between the plurality of patches by performing attention between the plurality of patches.

6. In paragraph 1, The step of determining whether there is an image of a region of interest with a lesion among the plurality of region of interest images using a third artificial intelligence model that inputs the plurality of embedding vectors is: A step of obtaining a lesion probability for each of the plurality of region-of-interest images using a third artificial intelligence model that takes the plurality of embedding vectors as input; A step of determining whether the above lesion probability exceeds a predefined first threshold value; A step of identifying an image of a region of interest in which the lesion probability exceeds the first threshold value as an image of a region of interest having the lesion, based on determining that the lesion probability exceeds the first threshold value; A method comprising a step of determining whether an image of a region of interest having the lesion exists among the plurality of images of regions of interest based on the identification result.

7. In paragraph 6, The step of determining whether there is an image of a region of interest with a lesion among the plurality of region of interest images using a third artificial intelligence model that inputs the plurality of embedding vectors is: A step of determining whether the lesion probability exceeds a predefined second threshold value based on determining that the lesion probability does not exceed the predefined first threshold value; A step of identifying an image of a region of interest in which the lesion probability exceeds the second threshold value as an image of a region of interest in which the lesion probability exceeds the second threshold value is determined to be an image of a region of interest in which the lesion probability cannot be confirmed; and A method comprising the step of identifying a region of interest image in which the lesion probability does not exceed the second threshold value as a region of interest image without a lesion, based on determining that the lesion probability does not exceed the second threshold value defined above.

8. In paragraph 1, A method wherein the plurality of signal values ​​include information about the substrate portion of each of the plurality of patches and information about the epithelial portion.

9. In paragraph 1, A method comprising the step of determining that the tissue sample is abnormal based on determining that an image of a region of interest having a lesion exists among the plurality of region of interest images.

10. In electronic devices, At least one processor comprising a processing circuit; and Containing at least one memory storing at least one instruction, By causing said at least one processor to execute said at least one instruction, said electronic device: Acquire the full image corresponding to the tissue sample, Acquire multiple region of interest images based on the entire image above, Obtaining multiple patches based on the multiple region of interest images, Using the first artificial intelligence model that takes the above multiple patches as input, multiple signal values ​​are extracted, Using a second artificial intelligence model that inputs the plurality of signal values ​​and the position values ​​corresponding to the plurality of patches, a plurality of embedding vectors each corresponding to each of the plurality of interest region images are obtained, Using a third artificial intelligence model that takes the plurality of embedding vectors as input, it is determined whether an image of a region of interest with a lesion exists among the plurality of region of interest images, An electronic device that determines that the tissue sample is normal based on determining that no region of interest image with a lesion exists among the plurality of region of interest images.

Images