Automatic tissue sketching system and sketching method for esophageal squamous cell carcinoma pathological image based on deep learning network

Abstract

The invention belongs to the field of biomedical information technology and artificial intelligence technology, and provides an automatic tissue sketching system and sketching method for esophageal squamous cell carcinoma pathological images based on a deep learning network, and the method comprises the steps: obtaining pathological images, and aiming at to-be-classified Hamp; e, carrying out data acquisition and image annotation on the image; preprocessing the acquired pathological image and the annotation result; acquiring a training image set and an independent verification image set, constructing seven tissue automatic sketching models of the pathological image by using transfer learning and RESNET50, and training and verifying the pathological image tissue automatic sketching network model based on deep learning; and inputting the preprocessed to-be-diagnosed pathological image into the target deep learning model to obtain an automatic tissue sketching classification result of each cutting region, endowing the automatic tissue sketching classification result with different colors for spatial reconstruction, and obtaining a tissue sketching result of the image. A visual result is provided, and a basis is provided for subsequent quantification of spatial tissue characteristics and clinical aid decision making.

Description

Technical Field

[0001] The present invention belongs to the fields of biomedical information technology and artificial intelligence technology, and specifically, relates to an automatic tissue delineation system and delineation method for esophageal squamous cell carcinoma pathological images based on a deep learning network. Background Art

[0002] Esophageal cancer is one of the most common digestive tract malignancies in the world. Its pathological types mainly include esophageal squamous cell carcinoma (ESCC) and esophageal adenocarcinoma (EADC). About 90% of esophageal cancer patients in my country suffer from esophageal squamous cell carcinoma. The cure rate of early esophageal squamous cell carcinoma exceeds 90%. However, due to the lack of early typical symptoms and the confirmation of biomarkers, most patients diagnosed with esophageal squamous cell carcinoma are in the T3-T4 stage, which results in a 5-year survival rate of only about 21%. At present, upper gastrointestinal endoscopy combined with histopathology is the gold standard for the diagnosis of esophageal squamous cell carcinoma. For lesions that are difficult to detect, they are mainly discovered by chromoendoscopy and electronic staining endoscopy, followed by targeted biopsy and diagnosis by histopathology.

[0003] Histopathological examination is to extract lesion specimens from the patient's organs or tissues, make pathological sections, observe the pathological changes and morphological changes under a microscope, and finally determine the type and severity of the disease. Pathological sections are usually stained with various colors using different dye substances to show different lesion information and help doctors diagnose diseases. For example, when diagnosing aortic aneurysm media degeneration, doctors often use a combination of five staining methods, including hematoxylin-eosin staining (H&E), ameba trophozoite staining (Alcian Blue and Periodic Acid-Schiff, AB-PAS), Masson Trichrome staining, smooth muscle actin (Actin, Smooth Muscle, SMA) staining and collagen fiber (Van Gieson, VG) staining, for observation and diagnosis. However, in clinical diagnosis, the use of multiple stained pathological images not only increases the workload of doctors, but also is time-consuming and labor-intensive, and often relies on the clinical experience of doctors, with the disadvantages of strong subjectivity, low accuracy, and poor repeatability. Therefore, designing and developing a computer-aided analysis system for various stained pathological images has clinical significance and practical value.

[0004] With the continuous improvement of computer computing speed and the improvement of algorithm design, the application of artificial intelligence, especially deep learning technology, in medical auxiliary diagnosis and clinical decision-making is becoming more and more mature. Computer-aided analysis systems based on deep learning can effectively process large-scale and high-precision pathological images, helping clinicians improve work efficiency and diagnostic accuracy. Summary of the invention

[0005] The present invention provides an automatic tissue delineation system and delineation method for esophageal squamous cell carcinoma pathological images based on a deep learning network. The method uses deep learning technology to construct a network model for automatic tissue delineation of multiple stained pathological images, and realizes an automatic delineation method for different tissues in multiple stained pathological images. The method provides auxiliary information for clinicians and improves the efficiency and accuracy of manual interpretation.

[0006] The purpose of the present invention is to provide a method for tissue delineation of esophageal squamous cell carcinoma pathological images, based on the pathological characteristics of different patients with esophageal squamous cell carcinoma on H&E images, using a deep convolutional neural network to achieve automatic delineation of esophageal squamous cell carcinoma pathological images. The results of automatic delineation can effectively utilize a large sample of pathological image data, reduce the non-subjective influence between sample data and the subjective assumptions of manual delineation, provide intuitive visualization results, and provide effective assistance for pathologists' diagnosis and film production quality control. Ultimately, the burden on pathologists is reduced and a basis is provided for subsequent quantitative spatial tissue features and clinical auxiliary decision-making.

[0007] The present invention is implemented by the following technical solution: an automatic tissue delineation method for esophageal squamous cell carcinoma pathological images based on a deep learning network, comprising the following steps: Obtain pathological images and manual delineation results: perform data collection and image annotation for H&E images to be classified; Preprocess the collected pathological images and annotation results, including staining standardization, cutting and color normalization; The training image set and the independent verification image set were obtained through the preprocessed annotation results. The automatic delineation model of seven tissues of pathological images was constructed by using transfer learning and deep convolutional neural network. The network model of automatic delineation of pathological image tissues based on deep learning was trained and verified. The preprocessed pathological image to be diagnosed is input into the target deep learning model to obtain the automatic tissue delineation and classification results of each cut area, and the automatic tissue delineation and classification results are assigned different colors for spatial reconstruction to obtain the tissue delineation results of the image.

[0008] Furthermore, the specific method for data collection and image annotation for the image to be classified is: screening out the pathological sections of the lesion location from the information system, using a scanning instrument to digitize the pathological sections, and manually annotating and image outlining a total of 7 different tissue types on the pathological image, wherein the 7 different tissue types are: connective tissue CON, squamous epithelium EPI, glandular GLA, lymphocyte LYM, smooth muscle MUS, cancer-related matrix STR and tumor TUM; Manually annotated tissue regions were segmented into non-overlapping image patches of 256 × 256 pixels (px) at 20x magnification to create a training set of 48,376 tissue sections from 40 H&E slides and a test set of 3996 tissue sections from 20 H&E slides; sections with glandular / squamous epithelial area >50% or tumor area <10% were considered poor quality and filtered out.

[0009] Furthermore, the method of preprocessing the acquired pathological images and the annotation results is as follows: using a high level of magnification, converting the large whole image into several small image patches of the same format; Constructing a pathological image dataset: aligning the acquired pathological images and corresponding annotations so that they have a consistent spatial relationship in the same coordinate system; Perform segmentation on the multiple stained pathology images of the annotated and registered images: Use Openslide to stitch all the images in the queue into non-overlapping 256×256 pixel (px) image patches at 20x magnification, and then perform color standardization; remove patches with a blank surface ratio of >50%; Color normalization based on Macenko's method: convert the RGB image to a specific color space; calculate the percentile of each channel in the color space and determine the threshold of the bright area; modify the image according to the threshold and map the color range to the standard range; Data augmentation operations were performed on various stained pathology images after cutting, segmenting and color normalization: the data augmentation operations included random transformations including random rotation within 10° and horizontal and vertical flipping.

[0010] The training image set and the independent verification image set are obtained through the preprocessed annotation results, wherein the training set and the test set do not contain image blocks from the same patient at the same time; The deep learning-based automatic pathology image delineation network model was trained and verified. The deep convolutional neural network RESNET50 model was used. The labeled pathology image dataset was used to supervise the RESNET50 model. The pathology images were input into the model, and the model automatically delineated the tissue according to the learned weights and feature representations. The training process used the back-propagation algorithm and optimizer to update the parameters. Use an independent validation image set to verify the training effect of the network model: use an independent test data set to evaluate the trained model, input the test images into the model, calculate the model's classification accuracy, recall rate, and precision, and evaluate its performance and accuracy of automatic tissue delineation.

[0011] The preprocessed pathological image to be diagnosed is input into the target deep learning model to obtain the automatic delineation result of each cut area, specifically including: using the trained pathological image automatic delineation network model to automatically classify the small pieces of tissue cut from the preprocessed pathological image, and organizing multiple pathological images into batches for processing; the target deep learning model will output the result of automatic tissue classification; After the automatic delineation is completed, the position information of each image patch is matched one by one with the automatic tissue classification result output by the model; all image patches are reassembled into a whole image to obtain the tissue delineation result, using different colors to highlight various tissue types; Among them: each image patch represents a small area in the original pathological image. After being processed and classified by the automatic outlining network model, it is assigned a specific tissue type label. By recombining these image patches according to their position information, an overall image can be generated, in which areas of different tissue types are highlighted in different colors.

[0012] In the present invention, H&E images were first annotated after collection, and several experienced pathologists manually annotated 60 H&E sections of ESCC tissues in the SXM-I cohort according to the same standards, and the image tissue types were divided into eight categories: background (BACK), connective tissue (CON), squamous epithelium (EPI), gland (GLA), lymphocyte (LYM), smooth muscle (MUS), cancer-associated stroma (STR), and tumor (TUM).

[0013] Manually annotated tissue regions were segmented into non-overlapping image patches of 256 × 256 pixels (px) at 20× magnification to create a training set of 48,376 tissue sections from 40 H&E slides and a test set of 3996 tissue sections from 20 H&E slides.

[0014] H&E slides with glandular / squamous epithelial area >50% or tumor area <10% will be considered poor quality and filtered out. A pathologist will review to further select high-quality and relevant images for analysis.

[0015] Digital images obtained from histological sections are usually multi-resolution by stitching together many small high-resolution images. Using high levels of magnification, such as 20x, the resulting images are often huge, making it difficult to read these sections using standard image analysis tools, graphics libraries, and deep neural networks.

[0016] Image segmentation can convert large whole slide images into several small image patches of the same format. All H&E slides in the queue were cut into non-overlapping 256×256 pixel (px) image patches at 20x magnification using Openslide (https: / / openslide.org / ), and then color normalized using the Macenko method. Patches with a high blank surface ratio (>50%) were removed.

[0017] Based on the aforementioned manually annotated tissue region to obtain a training set, the present invention also provides an image tissue automatic delineation system for esophageal cancer model training, characterized in that the system comprises: Image augmentation, which generates new samples by randomly transforming the images in the training dataset to expand the sample size; Construct a deep learning-based network model for automatic delineation of pathological image tissues and conduct network training; The deep learning-based network model for automatic delineation of pathological image tissue was verified.

[0018] The image enhancement operation includes: performing random rotation within 10 degrees and horizontal and vertical flipping of various stained pathology image patches, and combining random cropping to generate five times the image samples, so as to achieve the purpose of data expansion of the various stained pathology image datasets. The increase in the diversity of training samples can reduce the risk of overfitting and improve the robustness and generalization ability of the model.

[0019] In order to avoid classification bias, the training set and the test set should not contain image patches from the same patient at the same time.

[0020] Appropriate network structure and training method are crucial to classification performance. The RESNET50 model is a widely used deep convolutional neural network that was selected for training automatic tissue delineation due to its excellent classification performance.

[0021] To accelerate network convergence, we used transfer learning, that is, using the RESNET50 model pre-trained on the ImageNet dataset to initialize the network training. The training dataset was randomly taken from the public hand-drawn image queue NCT-CRC-HE-100k and data augmented. The internal validation set included 20% of the samples randomly selected from the NCT-CRC-HE-100k queue to produce an unbiased estimate of the model fit on the training dataset while adjusting the model hyperparameters.

[0022] The performance of the trained classifier was evaluated on an internal validation dataset and an independent test cohort CRC-VAL-HE-7K, consisting of 7180 image patches.

[0023] Furthermore, different colors are assigned to the automatic tissue delineation results for spatial reconstruction to obtain the tissue delineation results of the image.

[0024] After automatic tissue delineation, all image patches are reassembled into an overall slide image with different colors to highlight the various tissue types. This visualization technology provides a quick and convenient overview of the histological slide classification, which can help pathologists perform more effective quality control and diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the technical solutions of the embodiments disclosed in the present invention, the following briefly introduces the drawings required for describing the embodiments: Figure 1 A flowchart of the steps of a method and system for automatic tissue delineation of stained pathological images based on deep learning in the present invention; Figure 2 A schematic diagram of preprocessing the collected pathological images and annotation results in the present invention; Figure 3 For Figure 1 The block diagram shown is for training and validating a deep learning-based network model for automatic delineation of pathological image tissue; Figure 4 The figure is a schematic diagram of the automatic delineation result of a stained pathological image in an example of the present invention. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0027] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs, and the disclosure and materials cited therein are hereby incorporated by reference.

[0028] Technical equivalents to the specific embodiments described that are apparent to those skilled in the art using no more than routine experimentation are intended to be encompassed by this application.

[0029] The present invention relates to the fields of biomedical information technology, computer vision and image processing, and artificial intelligence technology, and in particular to a method and system for classifying multiple stained pathological images based on deep learning.

[0030] like Figure 1As shown, a flowchart of a method and system for automatic tissue delineation of stained pathological images based on deep learning includes the following steps: Step 101 : data collection and image annotation are performed for a variety of stained pathological images to be classified.

[0031] In order to train the ESCC tissue automatic delineation model, multiple pathological images with different stainings containing lesion locations were screened from the hospital information system. Experienced pathologists manually annotated 60 H&E sections of ESCC tissues in the SXM-I cohort according to the area and type of the lesions and the same standards described previously, and marked the corresponding positions on the multiple staining pathological images. Physicians used drawing software to annotate the images, and different colors indicated different lesion types. The types of manually annotated continuous pure tissue areas mainly included: connective tissue (CON), squamous epithelium (EPI), glands (GLA), lymphocytes (LYM), smooth muscle (MUS), cancer-associated stroma (STR), and tumors (TUM).

[0032] Step 102, preprocessing the collected pathological images and annotation results; Figure 2 As shown, digital images obtained from histological slides are often multi-resolution by stitching together many small high-resolution images. Using high levels of magnification, such as 20x, the resulting images are often huge. For example, in our in-house cohort SYSU-165, the average file size of H&E images is 2GB, with the largest size exceeding 5GB. Reading these slides is a huge challenge for standard image analysis tools, graphics libraries, and deep neural networks. Therefore, image segmentation is an important step that can convert large whole slide images into several small image patches of the same format.

[0033] Step 301, such as Figure 3 As shown, the present invention constructs a pathological image dataset. First, it relates to a position registration method for collected pathological images and annotations. In this method, the collected pathological images and corresponding annotations are firstly subjected to position registration operations. Position registration refers to aligning the spatial positions of the pathological images and annotations so that they have a consistent spatial relationship in the same coordinate system.

[0034] Then, the multi-stained pathology images of the annotated and registered images were segmented and color normalized. The segmentation used Openslide to stitch all H&E slides in the queue into non-overlapping 256×256 pixel (px) image patches at 20x magnification, and then color normalized. The patches with a high blank surface ratio (>50%) were removed.

[0035] In order to solve the influence of various technical factors on the appearance of histological images, the present invention proposes an appearance normalization method based on the Macenko method.

[0036] The Macenko method is a commonly used appearance normalization method that aims to eliminate color variations in histological images. This method is based on analyzing and adjusting the color space of histological images to counteract color differences caused by different production processes and scanning devices. The principle is to transform the color space of the image and map the color range of the image to a standard range to achieve normalization. The specific process is as follows: Convert RGB images to a specific color space, such as HED (Hematoxylin, Eosin, and DAB) space; Calculate percentiles for each channel in color space to determine thresholds for vivid areas; The image is corrected according to the threshold value to map the color range into the standard range.

[0037] Through appearance normalization based on the Macenko method, the color structure changes in histological images caused by the production and scanning process can be reduced. This helps to improve the deep neural network's ability to fit images, enabling it to more accurately learn and represent the structural features in the image, thereby providing more reliable results and accuracy for pathological image analysis, diagnosis and research.

[0038] Furthermore, data augmentation operations are performed on multiple stained pathology images after segmentation and color normalization. The data augmentation operations include random transformations including random rotation within 10 degrees and horizontal and vertical flipping. Increasing the diversity of training samples can reduce the risk of overfitting and improve the robustness and generalization ability of the model.

[0039] Step 103, obtaining a training set and an independent validation set through the preprocessed annotation results; In order to avoid classification bias, the training set and the test set do not contain image patches from the same patient at the same time.

[0040] Manually annotated tissue regions were segmented into non-overlapping tissue patches to form a training set of 48,376 tissue sections created from 40 H&E slides and a test set of 3996 tissue sections created from 20 H&E slides.

[0041] Among them, H&E slides with glandular / squamous epithelial area >50% or tumor area <10% will be considered as poor quality and filtered out. Pathologists will review to further select high-quality and relevant images for subsequent analysis.

[0042] Step 104, training and verifying the deep learning-based automatic delineation network model for pathological image tissue; In order to obtain good classification performance, it is crucial to select an appropriate network structure and training method. The present invention adopts the RESNET50 model, which is a widely used deep convolutional neural network, and its excellent classification performance makes it an ideal choice for training automatic tissue delineation.

[0043] During the training process, the RESNET50 model is supervised by using a labeled pathological image dataset. By inputting the pathological image into the model, the model automatically delineates the tissue based on the learned weights and feature representations. The training process uses the back propagation algorithm and optimizer to update parameters to minimize the loss function and improve the classification performance.

[0044] In order to verify the effect of training, an independent test data set can be used to evaluate the trained model. By inputting the test images into the model, the model's classification accuracy, recall rate, precision and other evaluation indicators can be calculated to evaluate the performance and accuracy of its automatic tissue delineation.

[0045] Step 302, such as Figure 3 As shown in the figure, it is the specific training process of automatically outlining the network model, which specifically includes the following steps: RESNET50 (Residual Network-50) is a deep convolutional neural network model architecture. It consists of 50 convolutional layers and fully connected layers and is widely used in image classification and target recognition tasks. The design concept of RESNET50 is to solve the gradient vanishing and gradient exploding problems in deep network training by introducing residual connections, thereby allowing the network to be deeper and easier to train.

[0046] Network training includes steps such as forward propagation, loss function, back propagation, optimizer, and iterative training. Through multiple iterations, the network weights and parameters are adjusted so that it can accurately predict tissue classification or delineation results.

[0047] To accelerate network convergence, we employed transfer learning, where we initialized the training of our classifier using a RESNET50 model pre-trained on the ImageNet dataset. The training dataset was randomly taken from a public hand-drawn image cohort, NCT-CRC-HE-100k, and data augmentation was performed. The internal validation set included 20% of the samples randomly selected from the NCT-CRC-HE-100k cohort to produce an unbiased estimate of the model fit on the training dataset while tuning the model hyperparameters.

[0048] Evaluation of network delineation: The performance of the trained classifier was evaluated in an internal validation dataset and an independent test cohort, CRC-VAL-HE-7K, which included 7,180 image patches. The model achieved high classification accuracy of 99% and 95% on the internal validation and test datasets, respectively, demonstrating the robustness of the classification model.

[0049] Step 105, inputting the preprocessed pathological image into the target model to obtain the automatic tissue classification result; First, preprocessed pathological image data is loaded from the database. These images are preprocessed, including segmentation, color normalization and other steps, to obtain image patches suitable for input into the network.

[0050] Since there are multiple pathological images that need to be input into the target model for processing, they are organized into batches for processing to improve computational efficiency. Through batch processing, multiple images can be passed to the target model at the same time, realizing the advantages of parallel computing and reasoning. This can not only significantly improve the computational efficiency of the model, but also speed up the generation of automatic tissue classification results.

[0051] Once the image data is organized into batches, the preprocessed image data is passed to the input layer of the target model and then input into a deep learning model based on a multi-layer convolutional neural network. The model will output the results of automatic organization and classification.

[0052] Step 106 , assigning different colors to the tissue classification results for spatial reconstruction to obtain tissue delineation results of the image.

[0053] like Figure 4 As shown in the figure, after the automatic tissue delineation is completed, a visualization technique is used to correspond the position information of each image patch to the automatic tissue classification results output by the model. By recombining all image patches into an overall slide image, different colors are used to highlight various tissue types. This visualization technique provides a quick and convenient overview of the histological section classification, which can help pathologists perform more effective quality control and diagnosis.

[0054] In this process, each image patch represents a small area in the original pathology image, and after being processed and classified by the automatic delineation network model, it is assigned a specific tissue type label. By recombining these image patches according to their position information, an overall slide image can be generated, in which areas of different tissue types are highlighted in different colors.

[0055] This visualization technology provides pathologists with an intuitive overview, allowing them to quickly understand the distribution of different tissue types throughout the tissue section. By using different colors to distinguish different tissue types, pathologists can more easily identify and analyze specific tissue structures in the section. This quick and intuitive overview helps speed up the diagnostic process and improve the work efficiency of pathologists.

[0056] Additionally, this visualization technology helps with quality control. By observing the distribution of different tissue types throughout the tissue slide, the pathologist can spot any anomalies or flaws during the diagnostic process. This helps ensure the quality of the tissue slide and identify any issues in a timely manner to ensure an accurate diagnosis.

[0057] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An automatic tissue delineation method for esophageal squamous cell carcinoma pathological images based on a deep learning network, characterized by: The steps include: Obtain pathological images and manual delineation results: perform data collection and image annotation for H&E images to be classified; Preprocess the collected pathological images and annotation results, including staining standardization, cutting and color normalization; The training image set and the independent verification image set were obtained through the preprocessed annotation results. The automatic delineation model of seven tissues of pathological images was constructed by using transfer learning and deep convolutional neural network. The network model of automatic delineation of pathological image tissues based on deep learning was trained and verified. The preprocessed pathological image to be diagnosed is input into the target deep learning model to obtain the automatic tissue delineation and classification results of each cut area, and the automatic tissue delineation and classification results are assigned different colors for spatial reconstruction to obtain the tissue delineation results of the image.

2. The method according to claim 1, characterized in that: The specific method for data collection and image annotation for the image to be classified is: screening out the pathological sections of the lesion location from the information system, using a scanning instrument to digitize the pathological sections, and manually annotating and image delineating and annotating a total of 7 different tissue types on the pathological image, wherein the 7 different tissue types are: connective tissue CON, squamous epithelium EPI, glandular GLA, lymphocyte LYM, smooth muscle MUS, cancer-related matrix STR and tumor TUM; Manually annotated tissue regions were segmented into non-overlapping image patches of 256 × 256 pixels at 20× magnification to create a training set of 48,376 tissue sections from 40 H&E slides and a test set of 3996 tissue sections from 20 H&E slides; sections with glandular / squamous epithelial area >50% or tumor area <10% were considered poor quality and filtered out.

3. The method according to claim 1, characterized in that: The method for preprocessing the collected pathological images and annotation results was as follows: all images in the queue were stitched into non-overlapping 256 × 256 pixel (px) image patches at 20x magnification using Openslide, and then color normalized; Plaques with a blank surface ratio of >50% were eliminated; Constructing a pathological image dataset: aligning the acquired pathological images and corresponding annotations so that they have a consistent spatial relationship in the same coordinate system; Color normalization based on Macenko method: converting RGB images to a specific color space; Calculate the percentile of each channel in the color space and determine the threshold of the bright area; modify the image according to the threshold and map the color range to the standard range; Data augmentation operations were performed on various stained pathology images after cutting, segmenting and color normalization: the data augmentation operations included random transformations including random rotation within 10° and horizontal and vertical flipping.

4. The method according to claim 1, characterized in that: The training image set and the independent verification image set are obtained through the preprocessed annotation results, wherein the training set and the test set do not contain image blocks from the same patient at the same time; The deep learning-based automatic delineation network model for pathological images was trained and verified. The deep convolutional neural network RESNET50 model was used. The RESNET50 model pre-trained on the ImageNet dataset was used to initialize the network training. The RESNET50 model was supervised for learning using a labeled pathological image dataset. The pathological images were input into the model, and the model automatically delineated tissues based on the learned weights and feature representations. The training process used the back-propagation algorithm and optimizer to update parameters. Use an independent validation image set to verify the training effect of the network model: use an independent test data set to evaluate the trained model, input the test images into the model, calculate the model's classification accuracy, recall rate, and precision, and evaluate its performance and accuracy of automatic tissue delineation.

5. The method according to claim 1, characterized in that: The preprocessed pathological image to be diagnosed is input into the target deep learning model to obtain the automatic delineation result of each cut area, specifically including: using the trained pathological image automatic delineation network model to automatically classify the small pieces of tissue cut from the preprocessed pathological image, and organizing multiple pathological images into batches for processing; the target deep learning model will output the result of automatic tissue classification; After the automatic delineation is completed, the position information of each image patch is matched one by one with the automatic tissue classification result output by the model; all image patches are reassembled into a whole image to obtain the tissue delineation result, using different colors to highlight various tissue types; Among them: each image patch represents a small area in the original pathological image. After being processed and classified by the automatic outlining network model, it is assigned a specific tissue type label. By recombining these image patches according to their position information, an overall image can be generated, in which areas of different tissue types are highlighted in different colors.

6. An automatic image tissue delineation system for esophageal squamous cell carcinoma pathological image training obtained by any method described in claims 1-5.

7. The automatic image tissue delineation system for esophageal squamous cell carcinoma pathological image training according to claim 6 is characterized by: The system includes: Image augmentation, which generates new samples by randomly transforming the images in the training dataset to expand the sample size; Construct a deep learning-based network model for automatic delineation of pathological image tissues and conduct network training; Verify the deep learning-based network model for automatic delineation of pathological image tissues; The image enhancement operation includes: performing random rotation within 10 degrees and horizontal and vertical flipping of the various stained pathological image patches, and combining random cropping to generate five times the image samples, so as to achieve data expansion of the various stained pathological image data sets; the training set and the test set do not contain image patches from the same patient at the same time; Use RESNET50 model to train automatic tissue delineation; The RESNET50 model pre-trained on the ImageNet dataset is used to initialize the network training; the training dataset is randomly obtained from the public hand-drawn image queue NCT-CRC-HE-100k and data augmentation is performed; the internal validation set includes 20% of the samples randomly selected from the NCT-CRC-HE-100k queue, while adjusting the model hyperparameters to produce an unbiased estimate of the model fit of the training dataset; The performance of the trained classifier was evaluated on an internal validation dataset and an independent test cohort, CRC-VAL-HE-7K, which includes 7180 image patches; Different colors are assigned to the automatic tissue delineation results for spatial reconstruction to obtain the tissue delineation results of the image; After automatic tissue delineation, all image patches were reassembled into an overall image with different colors used to highlight the various tissue types.

Images