Digitized pathological section-based digestive tract tumor micro-ecosystem analysis method

Abstract

The invention provides an analysis method of a digestive tract tumor micro-ecosystem based on a digital pathological section. The analysis method comprises the following steps: firstly, accurately extracting a foreground region of gastric cancer tissues by using an Otsu algorithm; then predicting cancer and non-cancer regions based on a convolutional neural network (CNN); a CRF algorithm is used to determine a junction area of the cancer area and the non-cancer area, and cell categories of the cancer area and the junction area are accurately predicted by means of a HoverNet model; and finally, quantitatively evaluating the survival and recurrence risk of the patient according to the characteristics extracted from the gastric cancer cells by applying a Cox proportional risk model and combining a lasso-cox method. The method has the advantages that the gastric cancer pathological image can be analyzed in an omnibearing and multi-level mode, and the technical bottlenecks that when a traditional algorithm is used for processing a complex pathological image, precision is insufficient, and judgment on the cancer area boundary and the cell category is not accurate are broken through; key data support can be provided for a doctor to formulate a personalized treatment scheme and accurate prognosis evaluation, and the scientificity and effectiveness of gastric cancer diagnosis and treatment are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of tumor micro-ecosystem, and particularly relates to a method for analyzing a digestive tract tumor micro-ecosystem based on a digital pathological section. BACKGROUND

[0002] 1. Tumor micro-ecosystem and its clinical significance:

[0003] Tumor represents a complex ecological system, which contains tumor cells, immune cells, cancer-associated fibroblasts (CAFs), endothelial cells (ECs), parietal cells, and various cells interacting with them, which together constitute a complex and mutually influencing tumor micro-ecosystem. The composition and functional state of the tumor micro-ecosystem can vary greatly depending on the type of tumor, intrinsic characteristics, tumor stage, and patient health status. In-depth understanding of the complex interactions between tumor cells and systemic mediators during disease progression is crucial for developing effective anti-cancer treatment plans.

[0004] Generally, the tumor micro-ecosystem includes numerous cells and non-cellular components. Among them, adaptive immune cells such as CD8+ T cells, CD4+ T cells, regulatory T cells (Tregs), and B cells play a key role in the anti-tumor process. Meanwhile, myeloid immune cells such as macrophages, neutrophils, monocytes, dendritic cells, mast cells, eosinophils, myeloid-derived myeloid suppressor cells (MDSCs), and platelets also widely participate in the pathological process of tumor occurrence and development.

[0005] In addition, tumor stromal cells and matrix components, such as cancer-associated fibroblasts (CAFs), adipocytes, extracellular matrix (ECM), neurons, and nerve fibers, including vascular cells such as vascular endothelial cells (ECs), lymphatic endothelial cells (LECs), and parietal cells, are also important components of this complex system.

[0006] The complexity of the tumor micro-ecosystem is reflected in the interactions of various components, such as cells in the tumor microenvironment interacting through direct contact and paracrine, etc., shaping a tumor microenvironment suitable for tumor growth or an anti-tumor microenvironment in different types and different patient tumor tissues. Understanding the composition of the tumor micro-ecosystem and the role of each component and its interaction in tumor occurrence and development is crucial for understanding the pathophysiological process of tumor, exploring therapeutic targets, and developing effective anti-cancer means.

[0007] 2. Limitations of previous methods for tumor micro-ecosystem:

[0008] Previous clinical practice and research on the characteristics of the gastrointestinal tumor microecosystem are mainly based on immunohistochemistry, multiple immunofluorescence, RNA-seq, scRNA-seq, and spatial transcriptomics, etc. The experimental operation procedures are complicated and the cost is high. Moreover, the simple transcriptome method cannot mark the spatial position information of each component in the tumor, so it is necessary to establish a more convenient and low-cost method to analyze the tumor microecosystem.

[0009] 3. Pathological algorithm and prognosis:

[0010] The existing pathological analysis technology of gastrointestinal tumors, especially gastric cancer, has many shortcomings, which seriously hinders the process of precise diagnosis and efficient treatment of gastric cancer. The main problems are in the two core blocks of pathological image recognition algorithm and prognosis analysis.

[0011] In the field of pathological image recognition algorithm, the problems are concentrated in the following key links:

[0012] 1) Foreground region extraction: In the past, it relied on manual segmentation or simple threshold segmentation algorithm. Manual segmentation is extremely inefficient, and because the experience and judgment standards of different operators are quite different, the results are highly subjective and lack consistency. The simple threshold segmentation algorithm has poor adaptability when faced with complex and variable gastric cancer pathological images. When encountering images with uneven lighting and strong background interference, it is easy to have incomplete foreground extraction, normal tissue is misjudged as background, or background is mistakenly segmented as foreground tissue, which seriously interferes with subsequent analysis.

[0013] 2) Cancer and non-cancer judgment and cancer area analysis: Traditionally, it mainly relies on the subjective visual judgment of pathologists. This process not only consumes time and effort, but also the experience levels of pathologists are uneven, and misdiagnosis and missed diagnosis occur from time to time. At the same time, cancer area analysis lacks effective image stitching technology, and cannot integrate multiple slices into a complete cancer area image. The quantitative analysis means is also seriously lacking, making it difficult to accurately obtain key information such as the shape, size and boundary of the cancer area, and further objectively evaluating the invasiveness of cancer cells, which undoubtedly brings great difficulties to the scientific formulation of subsequent treatment plans.

[0014] 3) Cell type recognition: It mainly relies on immunohistochemical staining combined with microscope observation, and the operation process is complicated, requiring multiple staining and washing steps, and the number of cell types that can be identified is limited. In the face of complex gastric cancer tissue, it is difficult to accurately distinguish cancer cells from inflammatory cells, interstitial cells, etc., which makes it difficult to deeply explore the role of different cells in the occurrence and development of gastric cancer, and is not conducive to a comprehensive understanding of the pathogenesis of gastric cancer.

[0015] In terms of prognosis analysis, there are also significant shortcomings:

[0016] 1) Feature extraction and analysis: In the past, only a few simple features such as cell size and nuclear-cytoplasmic ratio could be extracted for prognostic analysis, which could not fully reflect the biological characteristics of cells, such as cell metabolic activity, gene expression characteristics and other key information. Moreover, the feature extraction process lacks standardized procedures, and there are significant differences in sample collection, processing and measurement methods among different studies, making it difficult to compare and integrate data, which greatly hinders the exchange and promotion of research results.

[0017] 2) Survival and recurrence prediction: The traditional survival and recurrence prediction model based on limited features has unsatisfactory accuracy. Most models fail to fully consider the complex relationship between cell features and patient survival time and recurrence risk, and do not analyze molecular characteristics of tumors, individual differences of patients, treatment methods and other factors, lack of comprehensive consideration of multiple factors, and are difficult to provide accurate decision support for clinical treatment, which cannot meet the urgent needs of personalized medicine.

[0018] In summary, the existing pathological analysis technology of gastric cancer has obvious defects in each key link, and new technical solutions are urgently needed to break through these difficulties and improve the diagnosis and treatment level of gastric cancer, to bring better treatment effect and survival hope to the majority of patients. SUMMARY

[0019] In view of the above-mentioned shortcomings of the prior art, the purpose of the present application is to provide an analysis method of a gastrointestinal tumor microecosystem based on digital pathological sections, which is used to solve the problems that the existing pathological analysis technology of gastrointestinal tumors, especially gastric cancer, has many shortcomings in pathological image recognition algorithm and prognosis analysis, which seriously hinders the process of precise diagnosis and efficient treatment of gastric cancer.

[0020] To achieve the above objects and other related objects, the present application provides the following technical scheme:

[0021] An analysis method of a gastrointestinal tumor microecosystem based on digital pathological sections, the method comprising the following steps: obtaining a gastrointestinal tumor specimen to be analyzed, and performing sectioning and staining treatment on the gastrointestinal tumor specimen to be analyzed, obtaining a whole section image of the stained gastrointestinal tumor according to the sectioning and staining treatment result; extracting a foreground region of gastric cancer tissue from the whole section image of the stained gastrointestinal tumor, and performing sectioning treatment on the foreground region and then screening, obtaining a gastric cancer section image with appropriate area according to the screening result;

[0022] The area-appropriate gastric cancer slice image is input into the trained CNN model, the probability of each gastric cancer slice image belonging to cancer tissue or non-cancer tissue is obtained through the trained CNN model, and all slice images predicted as cancer tissue are restored according to the original spatial position of the slice image to obtain a complete cancer area image; the boundary area between the cancer area and the non-cancer area in the complete cancer area image is determined, and the cell types and distribution of the cancer area and the boundary area are determined, then morphological features, texture features and spatial features are comprehensively extracted from the gastric cancer cell image; a data set for survival recurrence analysis is constructed according to the morphological features, texture features and spatial features of the gastric cancer cells, the data set is evaluated by using a Cox proportional hazards model combined with a lasso-cox method, and the survival recurrence risk of the patient is predicted according to the evaluation result.

[0023] In an embodiment of the present application, the foreground region of gastric cancer tissue is extracted from the full slice image of the stained digestive tract tumor, comprising: using Otsu algorithm to extract the foreground region of gastric cancer tissue from the full slice image of the stained digestive tract tumor; the Otsu algorithm in this technical solution can not only accurately extract the foreground region, but also overcome the interference of light and background.

[0024] In an embodiment of the present application, the Otsu algorithm is used to extract the foreground region of gastric cancer tissue from the full slice image of the stained digestive tract tumor, comprising: the Otsu algorithm divides the pixels in the full slice image into two categories of foreground and background based on the difference in gray scale distribution, and calculates the variance between the two categories of foreground and background at each threshold value by traversing different gray scale threshold values, and extracts the foreground region of gastric cancer tissue from the full slice image of the stained digestive tract tumor according to the inter-class variance.

[0025] In an embodiment of the present application, before the area-appropriate gastric cancer slice image is input into the trained CNN model, and the probability of each gastric cancer slice image belonging to cancer tissue or non-cancer tissue is obtained through the trained CNN model, it further comprises: using a large number of labeled WSI slice data of gastric cancer and normal gastric tissue to train an initial CNN model, and obtaining a trained CNN model according to the training result, wherein in the training process, the initial CNN model automatically extracts and identifies the feature differences between gastric cancer tissue and normal tissue by learning the labeled data, and converts the feature differences into parameters and feature representations inside the model

[0026] In an embodiment of the present application, the determination of the boundary area between the cancer area and the non-cancer area in the complete cancer area image comprises: using a conditional random field algorithm to optimize all slice images predicted as cancer tissue, and determining the boundary area between the cancer area and the non-cancer area in the complete cancer area image.

[0027] In an embodiment of the present application, the use of the conditional random field algorithm optimizes all slice images predicted as cancerous tissue, and determines the boundary area between the cancerous area and the non-cancerous area in the complete cancerous area image, comprising: regarding each pixel in the complete cancerous area image as a node, regarding the spatial relationship between the pixels as the characteristics of the edges, adjusting the prediction results by calculating the probability of each node belonging to the cancerous area or the non-cancerous area, and combining the information of adjacent nodes, thereby obtaining the boundary area between the cancerous area and the non-cancerous area in the complete cancerous area image.

[0028] In an embodiment of the present application, the determination of the cell class and its distribution in the cancerous area and the boundary area comprises: inputting the cancerous area and the boundary area between the cancerous area and the non-cancerous area into the trained HoverNet model, and obtaining the cell class and its distribution in the cancerous area and the boundary area through the trained HoverNet model, wherein a large amount of WSI image data of gastric cancer and related tissues is used to train the HoverNet model, and the WSI image data is annotated in detail by professional pathologists, and the class information of each cell is accurately marked.

[0029] In an embodiment of the present application, the comprehensive extraction of morphological features, texture features and spatial features from the gastric cancer cell image comprises: the morphological features of the gastric cancer cells are mainly extracted based on the nuclear contour, the texture features of the gastric cancer cells are extracted through the gray level co-occurrence matrix, and the spatial features of the gastric cancer cells are based on the graph method to simulate the spatial relationship between cells.

[0030] In an embodiment of the present application, the use of the Cox proportional hazards model and the lasso-cox method to evaluate the data set, and the prediction of the survival recurrence risk of the patient according to the evaluation result comprises: using the Cox proportional hazards model to evaluate the relationship between the features in the data set and the survival time and recurrence risk of the gastric cancer patient, combining the lasso-cox method to screen out the most significant features affecting survival recurrence from the features in the data set, and performing weighted processing on the features, calculating the score value of each whole slice image according to the weighted processing result, and realizing the quantitative evaluation of the survival recurrence risk of the patient according to the score value.

[0031] As described above, the present application has the following beneficial effects:

[0032] The present invention uses the Otsu algorithm to accurately extract the foreground area of ​​gastric cancer tissue, predicts cancer and non-cancerous areas based on convolutional neural network (CNN), and then uses image stitching technology to restore the entire cancer area. The conditional random field (CRF) algorithm is then used to optimize the cancer area and determine the boundary area between the cancer area and the non-cancerous area. The HoverNet model is used to accurately predict the cell types of the cancer area and the boundary area, thereby being able to analyze gastric cancer pathology images in an all-round and multi-level manner, breaking through the technical bottleneck of traditional algorithms in processing complex pathology images with insufficient accuracy and inaccurate judgment of cancer area boundaries and cell types; and then systematically extracting morphological features and texture from gastric cancer cell images. The results show that the proposed method can effectively improve the prognosis of gastric cancer patients and the effectiveness of gastric cancer diagnosis and treatment, and can also effectively improve the prognosis of gastric cancer patients. The method can also effectively improve the prognosis of gastric cancer patients and the prognosis of gastric cancer patients. The method can effectively improve the prognosis of gastric cancer patients and the prognosis of gastric cancer patients. The method can effectively improve the prognosis of gastric cancer patients and the prognosis of gastric cancer patients. The method can effectively improve the prognosis of gastric cancer patients and the prognosis of gastric cancer patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Shown is a schematic diagram of the overall process of the method for analyzing the digestive tract tumor microecological system based on digital pathological sections disclosed in an embodiment of the present invention;

[0034] Figure 2 It shows a flow chart of a specific implementation of the method for analyzing the digestive tract tumor microecological system based on digital pathological sections disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0035] The following describes the embodiments of the present invention through specific examples. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. It should be noted that the following embodiments and features in the embodiments can be combined with each other unless there is a conflict.

[0036] The present invention relates to a method for analyzing the microecological system of digestive tract tumors based on digital pathological sections, which is a method for identifying and classifying the single cell morphology of different types of cells in digestive tract tumors through complete digital HE-stained tissue pathological sections, and analyzing the microecological system of digestive tract tumors. The process is as follows: Figure 1 As shown, the details are as follows:

[0037] At step 101, a digestive tract tumor specimen to be analyzed is obtained, and the digestive tract tumor specimen to be analyzed is subjected to slicing and staining treatment, and a full slice image of the stained digestive tract tumor is obtained according to a slicing and staining treatment result.

[0038] Specifically, the HE staining slice production process is as follows: in the research of digestive tract tumors, the pre-treatment of the specimen is crucial for subsequent pathological analysis, and the present application focuses on the processing flow of the digestive tract tumor specimen, especially the gastric cancer specimen. First, all the obtained digestive tract tumor specimens are quickly placed in a 4% neutral formalin solution for fixation. The 4% neutral formalin has good tissue fixation effect, can stabilize the morphology and structure of the cells, prevent cell autolysis and tissue degradation, and ensure the integrity and original characteristics of the specimen to be retained to the greatest extent.

[0039] The fixed specimen is subjected to paraffin embedding operation. Paraffin embedding is a commonly used tissue processing technique. By immersing the tissue in molten paraffin, the tissue is wrapped in a hard paraffin block after the paraffin cools and solidifies, providing a stable support structure for subsequent slicing. Then, the paraffin-embedded specimen is cut into a 4 μm thick slice using a slicing machine. This thickness has been scientifically verified to ensure the integrity of the slice and clearly display the cellular structure and morphological characteristics of the tissue under a microscope, facilitating subsequent staining and observation.

[0040] After slicing, HE staining is performed. HE staining is one of the most commonly used staining methods in histology and pathology. Hematoxylin can stain the cell nucleus blue, and eosin can stain the cytoplasm and extracellular matrix to different degrees of red. Through this contrast staining, the morphology, structure of cells and the hierarchy and composition of tissues can be clearly displayed, helping pathologists accurately identify different types of cells and tissue lesions. Finally, in order to facilitate data storage, transmission and remote analysis, the stained slice is digitally scanned and converted into a TIF format file, thereby obtaining a plurality of full slice images (WSI). The TIF format has high resolution, lossless compression and other characteristics, and can accurately save the image information of the slice, including the fine structure of the cells and the staining details, thereby providing high-quality image data basis for subsequent image-based pathological analysis and research.

[0041] At step 102, the foreground region of the gastric cancer tissue is extracted from the full slice image of the stained digestive tract tumor, and the foreground region is subjected to slicing and then screening, and a gastric cancer slice image with a proper area is obtained according to a screening result.

[0042] Specifically, 1, the specific steps of using Otsu algorithm to extract the foreground area of gastric cancer tissue are: in the pathological study of gastric cancer, the whole slice image (WSI) contains rich pathological information, but due to the influence of various factors in the acquisition process, such as noise interference of acquisition equipment, blank area existing in the periphery of tissue sample and different illumination conditions, etc., there are a large number of background interference in the image, which brings great challenge to the subsequent accurate analysis, therefore, accurate extraction of the foreground tissue area becomes the key first step; Otsu algorithm is based on the difference of gray scale distribution, and the pixels in the image are divided into foreground and background two categories, by traversing different gray scale thresholds, the variance between the foreground and background two categories under each threshold is calculated, the interclass variance is an important index to measure the difference degree between two categories of data, the greater the variance, the higher the discrimination degree between two categories of data; the goal of Otsu algorithm is to find a gray scale threshold that makes the interclass variance reach the maximum value, and the image is segmented by using this threshold, the pixels with gray scale value greater than the threshold are divided into foreground (i.e. gastric cancer tissue area), and the pixels with gray scale value less than the threshold are divided into background;

[0043] This adaptive threshold segmentation method based on Otsu algorithm has obvious advantages, it can automatically determine the threshold according to the gray scale distribution characteristics of the image itself, without manual intervention, and avoids the segmentation error caused by the subjectivity of manual threshold setting; at the same time, the algorithm has strong robustness to the illumination conditions and sample quality of the image, no matter whether the image is collected under strong light, weak light or uneven illumination conditions, or whether the sample quality is different, Otsu algorithm can accurately extract the foreground area of gastric cancer tissue, lay a solid foundation for subsequent accurate analysis, effectively avoid the analysis error caused by inaccurate foreground extraction, and ensure the reliability and accuracy of the subsequent analysis results.

[0044] 2, the specific steps of slicing and screening are: after successfully extracting the foreground area of gastric cancer tissue by using Otsu algorithm, considering the complexity of gastric cancer tissue and the convenience of subsequent analysis, the foreground area needs to be further sliced; the invention adopts fixed size slicing method to cut the foreground area into 256*256 pixel slices, this fixed size is considered from many aspects, on the one hand, the size of 256*256 pixel slice can ensure that each slice contains enough pathological information, which can reflect the local characteristics of gastric cancer tissue, such as the morphology and arrangement of cancer cells; on the other hand, this size achieves a good balance between computing resources and analysis efficiency, which is convenient for subsequent processing and analysis;

[0045] After slicing is completed, not all slices can provide effective information for subsequent analysis, based on this, it is necessary to screen the slices; the screening method based on area threshold is adopted in the present application, and the slices with the foreground area less than 30% are removed; because in the gastric cancer tissue, the effective pathological information contained in the slice with too small area is relatively scarce, it is difficult to comprehensively and accurately reflect the real characteristics of the cancer tissue, for example, the cancer cells in the gastric cancer tissue show specific distribution and infiltration mode, and the slice with too small area may not completely contain these key characteristics, thereby leading to misjudgment of important information such as distribution and infiltration of cancer cells; if these slices with too small area are included in the subsequent analysis process, not only unnecessary errors will be introduced, which will interfere with the accurate judgment of the gastric cancer condition, but also the calculation amount and analysis time will be increased, and the analysis efficiency will be reduced; by removing these small slices, the calculation amount in the subsequent analysis process can be greatly reduced, and the analysis efficiency can be improved; at the same time, it is ensured that the slices entering the subsequent process can effectively represent the characteristics of the gastric cancer tissue, and provide reliable data support for accurate diagnosis, the foreground area of these screened slices is larger, and more rich cancer cell information is contained, which can more accurately reflect the pathological characteristics of the gastric cancer tissue, and is helpful for doctors and researchers to make more in-depth and accurate analysis and diagnosis.

[0046] Step 103, input the gastric cancer slice image with appropriate area into the trained CNN model, obtain the probability that each gastric cancer slice image belongs to cancer tissue or non-cancer tissue through the trained CNN model, and restore all slice images predicted as cancer tissue according to the original spatial position of the slice image to obtain a complete cancer region image.

[0047] Specifically, the specific steps of predicting the cancer region and viewing the entire cancer region are as follows: in order to accurately judge the cancer and non-cancer regions in the gastric cancer slice and further understand the overall characteristics of the cancer region, the present application adopts a convolutional neural network (CNN) for prediction and analysis, and the CNN is a deep learning model specially designed for processing image data, which has strong feature extraction and pattern recognition capability;

[0048] Firstly, a large number of WSI slice data of labeled gastric cancer and normal gastric tissue are used to train the CNN model, and in the training process, the model automatically extracts and identifies the feature differences between the gastric cancer tissue and the normal tissue in texture, morphology and other aspects through learning these labeled data; for example, the morphology of gastric cancer cells is usually irregular, the nucleus is enlarged and deeply stained, the cytoplasm is out of proportion, and the cells are arranged in disorder; while the normal gastric tissue cells are regular in shape and arranged in order, the CNN model can learn these subtle feature differences and convert them into parameters and feature representations inside the model.

[0049] After the training is completed, the sliced and screened gastric cancer sections are input into the trained CNN model, the model predicts the sections according to the learned feature patterns, and outputs the probability that each section belongs to cancerous tissue or non-cancerous tissue; for the sections predicted as cancerous tissue, in order to comprehensively understand the morphology, size and boundary characteristics of the cancer area, image stitching technology is used to integrate these sections, the image stitching technology finds the overlapping areas between the sections by comparing the features such as texture, color, shape, etc. between adjacent sections, and accurately matches and stitches these overlapping areas, thereby stitching multiple cancerous tissue sections into a complete cancer area image;

[0050] From the stitched cancer area image, doctors and researchers can directly observe the overall morphology of the cancer area, judge whether the cancer area shape is regular, and whether the boundary is clear, and the cancer area shape irregular, the boundary fuzzy often prompts the cancer cells to have strong invasiveness, and may have infiltrated the surrounding tissues. The regular shape of the cancer area and the clear boundary may indicate that the invasiveness of the cancer cells is relatively weak; these information has important significance for preliminary evaluation of the invasiveness of cancer cells, and can provide key basis for subsequent diagnosis and treatment plan formulation; for example, for gastric cancer with strong invasiveness, more aggressive treatment methods may be needed, such as expanding the range of surgical resection, strengthening the intensity of chemotherapy or radiotherapy, etc. For gastric cancer with weak invasiveness, a relatively conservative treatment plan can be considered to reduce damage to the patient's body.

[0051] Step 104, determine the boundary area between the cancer area and the non-cancer area in the complete cancer area image, and determine the cell class and distribution of the cancer area and the boundary area, and then comprehensively extract morphological features, texture features and spatial features from the gastric cancer cell image.

[0052] Specifically, 1, the specific steps of using CRF algorithm to optimize the cancer area and find the boundary area between cancer and non-cancer are: in the pathological study of gastric cancer, accurately defining the boundary of the cancer area and determining the boundary area between cancer and non-cancer is crucial for in-depth understanding of the invasion mechanism of cancer cells and formulating precise treatment plans. However, due to the complexity of the cancer area boundary and the infiltration phenomenon of cancer cells, traditional prediction methods often fail to accurately complete this task, in order to solve this problem, the present application uses the conditional random field (CRF) algorithm to optimize the cancer area prediction result, and determines the boundary area between cancer and non-cancer.

[0053] The CRF algorithm is a machine learning algorithm based on a probabilistic graph model that can fully consider the spatial relationships and contextual information between pixels in an image. In gastric cancer research, the CRF algorithm constructs a probabilistic model based on pixel features and the relationships between adjacent pixels. Specifically, each pixel is treated as a node, and the spatial relationships between pixels (such as Euclidean distance, gray difference, color similarity, etc.) are used as edge features. By calculating the probability of each node belonging to the cancerous or non-cancerous region and adjusting the prediction results in combination with the information of adjacent nodes, the cancerous region boundary is made more accurate and smooth. The joint probability distribution of the CRF model is determined by the following formula: where the unary potential function characterizes the relationship between a single node feature and the label, and the binary potential function describes the relationship between adjacent nodes. For a single node v, the probability of belonging to label y v is obtained by marginalizing the labels of other nodes, i.e. The maximum a posteriori probability estimate is commonly used in combination with a dynamic programming algorithm to solve it. After obtaining the probability of each node, the prediction results need to be adjusted in combination with the information of adjacent nodes. The belief propagation algorithm is commonly used. Initially, for each edge (u, v) ∈ E, the message m u→v (y v ) from node u is set to 1. When updating the message, it is based on the formula:

[0054] In determining the boundary region between cancer and non-cancer, the CRF algorithm uses its powerful modeling ability for pixel relationships to accurately identify pixels in the transition zone between cancerous and non-cancerous regions. The cell characteristics in the boundary region of gastric cancer are complex, with cancer cells and normal cells mixed together, and the cell morphology and structure gradually changing. The CRF algorithm can accurately divide this boundary region by considering the gray value, texture features, and association with surrounding pixels of the pixels. For example, in the boundary region, cancer cells may have an impact on surrounding normal cells, causing changes in the morphology, texture, and gene expression of normal cells. The CRF algorithm can capture these subtle changes to accurately determine the range and boundary of the boundary region. Accurate division of the boundary region between cancer and non-cancer provides an important basis for in-depth study of the invasion mechanism of cancer cells. By studying the cell characteristics, molecular markers, and cell-cell interactions in the boundary region, it can be revealed how cancer cells break through the defenses of normal tissues and infiltrate and spread to surrounding tissues, which has important guiding significance for the development of new anticancer drugs and treatment strategies, and helps to improve the treatment effect of gastric cancer and the survival rate of patients.

[0055] 2. The specific steps of using the HoverNet model to predict the six cell classes of the cancerous area and the border area are: in order to deeply understand the occurrence and development mechanism of gastric cancer, it is crucial to accurately identify the cell classes and their distribution in the cancerous area and the border area; the trained HoverNet model is used to classify the cells in the cancerous area and the border area of gastric cancer, and the cells are divided into six types: cancer cells (T), inflammatory cells (I), interstitial cells (S), necrotic cells (NC), normal cells (N) and unlabeled cells (No-label);

[0056] The HoverNet model is based on a convolutional neural network architecture, which combines semantic segmentation and instance segmentation techniques. Semantic segmentation can divide different regions in an image into different categories, while instance segmentation can further distinguish different instances within each category based on semantic segmentation. During the training of the HoverNet model, a large number of WSI image data of gastric cancer and related tissues were used, which were annotated in detail by professional pathologists, accurately marking the class information of each type of cell. Through learning from these labeled data, the HoverNet model can accurately identify the six types of cells and their distribution in the cancerous area and the border area. Different cells play different roles in the development of gastric cancer, for example, cancer cells are the core pathogenic factors of gastric cancer, and their continuous proliferation and invasion lead to tumor growth and spread. Inflammatory cells participate in the body's immune response, trying to resist cancer cells, but in some cases may also promote tumor development. Interstitial cells provide a growth microenvironment for cancer cells, interact with cancer cells, and affect the growth, migration and invasion ability of cancer cells. Necrotic cells are the products of cell death, and their presence may reflect internal hypoxia, nutrient deficiency, etc. Normal cells serve as a control to help researchers understand the differences between cancer cells and normal cells.

[0057] Accurate classification of cells helps to deeply understand the occurrence and development mechanism of gastric cancer, providing key information for diagnosis and treatment. For example, by analyzing the distribution ratio and spatial relationship of different cell types in the cancerous area and the border area, the malignancy, invasion ability of the tumor and the prognosis of the patient can be evaluated. This has important guiding significance for doctors to develop personalized treatment plans, helping doctors choose the most suitable treatment method for patients, improving treatment effectiveness and patient quality of life.

[0058] 3. The specific steps of extracting various features of cells - morphological features, texture features, and spatial features are: based on the accurate prediction of cell classes in the previous step, morphological, texture and spatial features are comprehensively extracted from the gastric cancer cell images, which are of great significance for in-depth study of the biological characteristics of gastric cancer cells and diagnosis of gastric cancer.

[0059] The morphological features are mainly based on the extraction of nuclear contours, including multiple key indicators; cell area is a basic morphological feature, which reflects the size of the cell; in the study of gastric cancer, cancer cells usually have larger cell areas than normal cells due to abnormal proliferation; by measuring the cell area, it can be preliminarily judged whether the cell has canceration; the centrifugal rate describes the flatness of the cell shape, cancer cells often show irregular shapes due to abnormal internal structure and metabolism, and the centrifugal rate is significantly different from that of normal cells; the roundness measures the degree of cell approaching to a circle, the morphology of normal cells is relatively regular, and the roundness is high, while the morphology of cancer cells is abnormal, and the roundness changes; in addition, it also includes features such as major axis length, minor axis length and perimeter, which reflect the size of the cell in different directions, and the perimeter describes the contour length of the cell. These morphological features can comprehensively and comprehensively depict the morphology of the cell, and provide a strong basis for distinguishing cell types and judging the health status of the cell;

[0060] Texture features are extracted by gray level co-occurrence matrix; gray level co-occurrence matrix is a statistical method for describing the distribution of gray values in an image, which can reflect the spatial correlation and gray variation of pixels in the image; from the gray level co-occurrence matrix, multiple texture features can be extracted, such as corner metric, which reflects the uniformity and regularity of the image gray distribution; in gastric cancer cells, due to the abnormal internal structure and metabolism, texture contrast, entropy and other features are significantly different from normal cells, for example, the enlargement of the nucleus and the increase of chromatin in cancer cells lead to more complex internal gray distribution, increased texture contrast and changed entropy value; these texture features reveal the state and structure information of the cell from a microscopic level, which helps to further distinguish cancer cells and normal cells and understand the biological characteristics of gastric cancer cells;

[0061] Spatial features are based on the graph-based method to simulate the spatial relationship between cells; cells are regarded as nodes in the graph, and the connection relationship between cells is regarded as edges, the spatial distribution of cells is described by constructing the intercellular graph, for example, the minimum edge length and the average edge length of the intercellular graph reflect the distance between cells, the change of the distance may be related to the invasion and metastasis ability of cancer cells; the degree number represents the number of connections between each cell and other cells, which reflects the activity and interaction intensity of the cell in the group, the core nature describes the position importance of the cell in the group, some cells located in the key position may play a key role in the development of gastric cancer; by analyzing these spatial features, the distribution pattern and interaction mechanism of gastric cancer cells in the tissue can be studied, which provides important clues for understanding the occurrence and development of gastric cancer.

[0062] In step 105, a data set for survival recurrence analysis is constructed according to the morphological features, texture features and spatial features of the gastric cancer cells, the data set is evaluated using the Cox proportional hazards model combined with the lasso-cox method, and the survival recurrence risk of the patient is predicted according to the evaluation result.

[0063] Specifically, the specific steps of survival recurrence analysis according to cell features are as follows: morphological, texture and spatial features of cells, and overall structural features of tissues are extracted from the staining image to construct a data set for survival recurrence analysis. These features contain rich pathological information and can reflect the severity of the disease and the prognosis of the gastric cancer patient.

[0064] The Cox proportional hazards model is used to evaluate the relationship between these features and the survival time and recurrence risk of the gastric cancer patient. The Cox proportional hazards model is a commonly used survival analysis method, which can analyze the influence of multiple covariates (i.e. extracted cell and tissue features) on survival time and recurrence risk. The feature data of the patient is taken as input, and the survival time and recurrence are taken as output. Through training on a large amount of patient data, the model can find features that have a significant impact on survival recurrence. According to the best result and weight of lasso-cox, the score value of each WSI image is calculated. Lasso-cox is a method combining lasso (least absolute shrinkage and selection operator) and Cox proportional hazards model, which can screen out the most significant features affecting survival recurrence from numerous features and weight these features. By calculating the score value of each WSI image, the survival recurrence risk of the gastric cancer patient can be quantitatively evaluated. The higher the score value, the shorter the survival time of the patient and the higher the recurrence risk. Conversely, the lower the score value, the better the survival prognosis of the patient. This quantitative evaluation result can provide an important reference for doctors to develop individualized treatment plans and prognosis evaluation, which helps to improve the treatment effect of gastric cancer and the survival rate of patients.

[0065] In practical applications, the research summary of the implementation case is as follows: the process used by the application: more than 300 cases of patients who underwent radical gastrectomy for gastric malignant tumor in Zhongshan Hospital Affiliated to Fudan University from the beginning of 2009 to 2018 and whose postoperative pathology was confirmed as gastric adenocarcinoma were retrospectively collected. The patient inclusion criteria are as follows: no other malignant tumor, complete clinical pathological image data and follow-up information, no preoperative chemotherapy, etc. The clinical basic information of the above-mentioned patients is recorded, such as age, gender, tumor characteristics (such as number, size and location, etc.), pathological characteristics (such as tumor differentiation, invasion depth, vascular invasion, lymph node metastasis, MSI, Her2 status, EBV status, TPS score, etc.), blood biochemical indicators (such as blood routine, liver and kidney function, electrolyte, tumor marker) and the like. The patients are regularly followed up, such as postoperative recovery, whether to receive adjuvant chemotherapy, which chemotherapy regimen to receive and chemotherapy cycle, whether to relapse after operation, relapse time, relapse site and treatment regimen after relapse, efficacy of relapse treatment, postoperative survival state, death time and death cause and the like. On the basis of the preliminary preparation work, an AI-based recognition model (such as shown in the formula (I)) is established: the WSI data set is randomly divided into 8:1:1, the foreground region is extracted by using the Otsu algorithm, the foreground region is segmented into 256*256 pixel slices, and preliminary screening is performed; the Transformer model is trained and the cancer and non-cancer are predicted, and the CRF algorithm is used to draw the boundary region; the HoverNet model trained is used to identify the cells in each region, and the morphology, texture and spatial features of each category of cells are extracted. And further combined with clinical data, the disease survival and recurrence risk are calculated and evaluated. Figure 2

[0066] In summary, the pathological image recognition algorithm principle and prognosis analysis principle of the application are as follows:

[0067] 1. Pathological image recognition algorithm principle:

[0068] 1) Foreground region extraction: the Otsu algorithm is adopted. The Otsu algorithm automatically determines the optimal gray threshold value by maximizing the inter-class variance according to the image gray value distribution characteristics, so as to realize accurate segmentation of foreground and background. By taking advantage of the difference in gray distribution between gastric cancer tissue and background, not only can the problems of uneven illumination and background interference be effectively overcome, but also the accuracy and stability of foreground region extraction can be improved.

[0069] ​2) Cancer and non-cancer judgment and cancer area analysis: Using the powerful feature extraction and pattern recognition ability of convolutional neural network (CNN), a large number of labeled gastric cancer and normal gastric tissue section data are used for training, so that the model learns the feature differences of the two in texture, morphology, etc., thereby accurately predicting cancer and non-cancer areas; for cancer area analysis, image stitching technology is used, based on the feature matching of adjacent sections, to restore the entire cancer area image, realize the quantitative analysis of the morphology, size and boundary of the cancer area, and provide objective basis for evaluating the invasiveness of cancer cells.

[0070] 3) Cell type recognition: With the help of HoverNet model based on convolutional neural network architecture and combining semantic segmentation and instance segmentation technology, a large number of detailed labeled gastric cancer and related tissue image data are used for training, so that the model can accurately identify six cell types such as cancer cells, inflammatory cells and interstitial cells and their distribution in cancer area and boundary area, and further reveal the mechanism of different cells in the development of gastric cancer.

[0071] 2) Prognosis analysis principle:

[0072] 1) Feature extraction and analysis: Comprehensive extraction of cell morphological, textural and spatial features to build a rich feature dataset. Morphological features such as cell area and centrifugal rate reflect cell characteristics in terms of shape and size; texture features are extracted through gray level co-occurrence matrix to reveal the microstructure of cells; spatial features simulate the spatial relationship between cells based on graph method, which comprehensively reflects the biological characteristics of cells; at the same time, standardized feature extraction process is developed to ensure data consistency and comparability.

[0073] 2) Survival and recurrence prediction: Cox proportional hazards model is used to evaluate the relationship between features and gastric cancer patient survival time and recurrence risk, and lasso-cox method is used to screen out features that significantly affect survival and recurrence and calculate their weights, to obtain the score value of each WSI image, and realize the quantitative evaluation of patient survival and recurrence risk, and provide precise decision support for clinical treatment.

[0074] The present application has the following advantages: 1. Pathological image recognition is more accurate and efficient: the constructed pathological image recognition algorithm system covers various advanced algorithms, such as Otsu algorithm for accurately extracting foreground area, which can overcome light and background interference; CNN predicts cancer and non-cancer area and displays the whole picture of cancer area through image splicing, CRF algorithm optimizes the boundary of cancer area and determines the junction area, HoverNet model accurately identifies six kinds of cell categories; these algorithms work together to comprehensively improve the analysis ability of gastric cancer pathological image, break through the bottleneck of traditional algorithm in accuracy, cancer area boundary judgment and cell category recognition, and provide more reliable basis for pathological diagnosis; 2. Prognosis analysis is more scientific and effective: the morphological, textural and spatial features of cells are systematically extracted, a rich data set is constructed, and the biological characteristics of cells are fully reflected; Cox proportional hazards model is used combined with lasso-cox method for analysis, and the score of WSI image is calculated to quantitatively evaluate the survival and recurrence risk of patients, changing the previous situation of lack of accurate quantitative indicators in prognosis evaluation, providing key data support for doctors to develop individualized treatment plan, and helping to improve treatment effect and patient survival rate.

[0075] The above examples only exemplarily illustrate the principles and effects of the present application, and are not used to limit the present application. All equivalent modifications or changes made by those skilled in the art without departing from the spirit and technical thought disclosed by the present application should still be covered by the claims of the present application.

Claims

1. A method for analyzing the digestive tract tumor microecology system based on digital pathological sections, characterized in that: The method comprises the following steps: Obtaining a digestive tract tumor specimen to be analyzed, and performing a section staining process on the digestive tract tumor specimen to be analyzed, and obtaining a whole section image of the stained digestive tract tumor according to the section staining process result; Extracting a foreground area of ​​gastric cancer tissue from the stained whole-slice image of the digestive tract tumor, slicing the foreground area and then screening it, and obtaining a gastric cancer slice image with a suitable area according to the screening result; Inputting gastric cancer slice images of appropriate area into a trained CNN model, obtaining the probability of each gastric cancer slice image belonging to cancerous tissue or non-cancerous tissue through the trained CNN model, and restoring all slice images predicted to be cancerous tissue according to their original spatial positions to obtain a complete cancer area image; determining the boundary region between the cancerous region and the non-cancerous region in the complete cancerous region image, determining the cell types and their distribution in the cancerous region and the boundary region, and then comprehensively extracting morphological features, texture features, and spatial features from the gastric cancer cell image; A data set for survival and recurrence analysis was constructed based on the morphological, texture, and spatial characteristics of the gastric cancer cells. The data set was evaluated using a Cox proportional hazards model combined with the lasso-cox method, and the patient's survival and recurrence risk was predicted based on the evaluation results.

2. The method for analyzing the digestive tract tumor microecology system based on digital pathological sections according to claim 1, characterized in that: The step of extracting the foreground area of ​​the gastric cancer tissue from the stained whole-slice image of the digestive tract tumor comprises: The Otsu algorithm was used to extract the foreground area of ​​gastric cancer tissue from the stained whole-slice images of digestive tract tumors.

3. The method for analyzing the digestive tract tumor microecology system based on digital pathological sections according to claim 2, characterized in that: The method of extracting the foreground area of ​​gastric cancer tissue from the stained whole-slice image of the digestive tract tumor using the Otsu algorithm includes: The Otsu algorithm divides the pixels in the whole-slice image into two categories: foreground and background based on grayscale distribution differences, and calculates the variance between the foreground and background categories at each threshold by traversing different grayscale thresholds. Based on the inter-class variance, the foreground area of ​​gastric cancer tissue is extracted from the stained whole-slice image of the digestive tract tumor.

4. The method for analyzing the digestive tract tumor microecology system based on digital pathological sections according to claim 1, characterized in that: Before inputting the gastric cancer slice images of appropriate areas into the trained CNN model and obtaining the probability of each gastric cancer slice image belonging to cancerous tissue or non-cancerous tissue through the trained CNN model, the method further includes: The initial CNN model was trained using a large amount of labeled WSI slice data of gastric cancer and normal gastric tissue, and a trained CNN model was obtained based on the training results. During the training process, the initial CNN model automatically extracted and identified the characteristic differences between gastric cancer tissue and normal tissue by learning from the labeled data, and converted the characteristic differences into parameters and feature representations within the model.

5. The method for analyzing the digestive tract tumor microecology system based on digital pathological sections according to claim 1, characterized in that: The determining of the boundary region between the cancerous area and the non-cancerous area in the complete cancerous area image includes: A conditional random field algorithm is used to optimize all slice images predicted to be cancerous tissues, and a boundary area between the cancerous area and the non-cancerous area in the complete cancerous area image is determined.

6. The method for analyzing the digestive tract tumor microecology system based on digital pathological sections according to claim 5, characterized in that: The method of using the conditional random field algorithm to optimize all slice images predicted to be cancerous tissues and determining the boundary region between the cancerous area and the non-cancerous area in the complete cancerous area image comprises: Each pixel in the complete cancerous area image is regarded as a node, and the spatial relationship between pixels is used as the feature of the edge. By calculating the probability of each node belonging to the cancerous area or the non-cancerous area and adjusting the prediction result based on the information of adjacent nodes, the boundary area between the cancerous area and the non-cancerous area in the complete cancerous area image is obtained.

7. The method for analyzing the digestive tract tumor microecology system based on digital pathological sections according to claim 1, characterized in that: Determining the cell types and distribution in the cancerous area and the border area includes: The cancerous areas and the boundary areas between the cancerous areas and non-cancerous areas were input into a trained HoverNet model, and the cell types and their distribution in the cancerous areas and the boundary areas were obtained using the trained HoverNet model. A large amount of WSI image data of gastric cancer and related tissues was used when training the HoverNet model. The WSI image data was carefully annotated by professional pathologists to accurately mark the category information of each cell.

8. The method for analyzing the digestive tract tumor microecology system based on digital pathological sections according to claim 1, characterized in that: The comprehensive extraction of morphological features, texture features, and spatial features from gastric cancer cell images includes: The morphological features of the gastric cancer cells are mainly extracted based on nuclear contours, the texture features of the gastric cancer cells are extracted through a gray-level co-occurrence matrix, and the spatial features of the gastric cancer cells are simulated based on a graph-based method for spatial relationships between cells.

9. The method for analyzing the digestive tract tumor microecology system based on digital pathological sections according to claim 1, characterized in that: The Cox proportional hazard model is used in combination with the lasso-cox method to evaluate the data set, and the patient's survival and recurrence risk is predicted based on the evaluation results, including: The Cox proportional hazards model was used to evaluate the relationship between the features in the dataset and the survival time and recurrence risk of gastric cancer patients. The lasso-Cox method was combined to screen out the features in the dataset that had the most significant impact on survival and recurrence. The features were weighted, and the score value of each whole-slice image was calculated based on the weighted processing results. The score value was used to quantitatively assess the patient's survival and recurrence risk.

Images