A deep learning system and construction method for colorectal adenoma identification and semantic segmentation

Abstract

The present invention belongs to the technical field of learning model construction, and provides a semantic segmentation deep learning system and construction method for colorectal adenoma identification, including: a data set acquisition and processing module, which is used to acquire and process case images of colorectal adenoma to obtain a target data set; a model construction module, which is used to construct a semantic segmentation deep learning model that obtains H&E adversarial average teacher based on a dye called hematoxylin and eosin in combination with a semi-supervised learning model; and a model learning and training module, which is used to learn and train the semantic segmentation deep learning model based on the target data set. The present invention can ensure the creation quality of the semantic segmentation deep learning system for colorectal adenoma identification by acquiring and processing case images of colorectal adenoma, and based on a dye called hematoxylin and eosin in combination with a semi-supervised learning model, construct and learn and train a semantic segmentation deep learning model, thereby providing a basis for efficiently and accurately identifying case images of colorectal adenoma.

Description

Technical Field

[0001] The present invention relates to the technical field of learning model construction, and in particular to a colorectal adenoma recognition semantic segmentation deep learning system and a construction method. Background Art

[0002] Histopathology provides the most accurate clinical information for tumor diagnosis. However, manual evaluation can be time-consuming and subjective, even for experienced pathologists. This has led to the development of image analysis algorithms. In particular, deep learning-based methods have played a significant role in the field of computational pathology in recent years.

[0003] While deep convolutional neural networks have demonstrated remarkable performance, domain conversion between training data (source domain) and test data (target domain) often hinders their generalization capabilities in real-world applications. In computational pathology, variations in staining color are an important factor contributing to staining; hematoxylin and eosin (H&E), a popular stain applied to the entire slide image, renders cell nuclei blue (hematoxylin) and cytoplasmic regions pink (eosin). Multiple factors involved in the slide preparation, staining, and scanning processes affect the final color appearance of case images of colorectal adenomas, such as sample thickness, tissue fixation, staining protocols, scanning devices, and lighting conditions. As a result, slides from different medical institutions often exhibit different visual characteristics, often causing the performance of deep convolutional neural networks to drop dramatically when deployed to a new hospital.

[0004] To address this issue, existing methods aim to perform color normalization of dyeing, typically using image processing methods and deep generative models to segment the color distribution of training and test data; and focus on simulating color changes of dyeing through color enhancement techniques; however, modifying the features of the input image does not always guarantee correct conversion and may introduce additional noise to the training phase.

[0005] Therefore, it is necessary to provide a semantic segmentation deep learning system and construction method for colorectal adenoma identification. Summary of the Invention

[0006] The present invention provides a semantic segmentation deep learning system and construction method for colorectal adenoma identification. By acquiring and processing case images of colorectal adenomas, and based on a dye called hematoxylin and eosin (H&E), combined with a semi-supervised learning model, a semantic segmentation deep learning model of H&E Adversarial Average Teacher (HEAMT) is constructed, and the semantic segmentation deep model is trained. This can ensure the creation quality of the semantic segmentation deep learning system for colorectal adenoma identification, and provide a basis for efficiently and accurately identifying case images of colorectal adenomas.

[0007] The present invention provides a colorectal adenoma recognition semantic segmentation deep learning system, comprising:

[0008] The data set acquisition and processing module is used to acquire and process case images of colorectal adenoma to obtain the target data set;

[0009] A model building module for building a semantic segmentation deep learning model based on a dye called hematoxylin and eosin (H&E) combined with a semi-supervised learning model (Adversarial Mean Teacher, AMT);

[0010] The model learning and training module is used to learn and train the semantic segmentation deep model based on the target dataset.

[0011] Furthermore, the data set acquisition and processing module includes a data set acquisition unit and a data set processing unit;

[0012] A data set acquisition unit is used to acquire case images of colorectal adenoma to form an original data set;

[0013] The data set processing unit is used to process the original data set to obtain the target data set.

[0014] Furthermore, the original data set is processed to obtain the target data set, including:

[0015] Divide the original data set into training data set and test data set;

[0016] Processing the training data set to obtain a processed training data set; wherein the processing includes color normalization processing and color enhancement using the HSV perturbation method; the color normalization processing is performed to make it have a color distribution similar to that of the test data set; the color enhancement is performed to expand the training data set;

[0017] Aggregate the test dataset and the processed training dataset to form the target dataset.

[0018] Furthermore, based on the stain called hematoxylin and eosin (H&E), combined with a semi-supervised learning model (Adversarial Mean Teacher, AMT), a semantic segmentation deep learning model for obtaining H&E Adversarial Mean Teacher (HEAMT) was constructed, including:

[0019] Get the teacher (Mean Teacher, MT) model and the student model;

[0020] Combine the student model and the teacher model to obtain a deep learning model; the teacher model is used to generate high-quality pseudo labels to supervise the student model's learning in the target domain;

[0021] In the deep learning model, a H&E discriminator with a gradient reversal layer (GRL) is introduced to perform H&E-oriented adversarial learning, generating a semantic segmentation deep learning model with the H&E Adversarial Average Teacher (HEAMT). Among them, adversarial learning is used to guide the semantic segmentation deep learning model to learn coloring-invariant representation through a min-max game.

[0022] Furthermore, the semantic segmentation deep model is trained, including:

[0023] Set the training dataset S and test dataset T in the target dataset; where S includes N s Case images of labeled colorectal adenomas, in, Represents the i-th sample data in the training data set S; T includes N t Case images of unlabeled colorectal adenomas; in, Represents the i-th sample data in the training data set T;

[0024] According to the divided stages, based on the training dataset S and the test dataset T, the loss is optimized in an end-to-end manner to perform learning and training; the stages include the initialization stage and the mutual learning stage;

[0025] The initialization phase involves training a fully supervised classification model based on the training dataset S, i.e., training the classifier using labeled colorectal adenoma case images. The classifier is then replicated into two copies, which are used as the initialized student model and teacher model, respectively. The mutual learning phase involves the teacher model obtaining weakly enhanced images from the test dataset T, and the student model obtaining strongly enhanced images from the training dataset S and the test dataset T. The teacher model generates pseudo labels on the test dataset T to train the student model, which then updates the teacher model using an exponential moving average.

[0026] Using the calculated H&E matrix, an aH&E discriminator with a gradient reversal layer (GRL) is used to align the staining distributions in the feature space.

[0027] Furthermore, the mutual learning phase also includes:

[0028] Using strong and weak enhancement strategies, the student model is helped to learn coloring invariant features in the training dataset S and the test dataset T;

[0029] Inverse learning is introduced into the semantic segmentation deep learning model to align the distribution of the training dataset S and the test dataset T in the feature space.

[0030] Furthermore, the weak enhancement strategy includes random horizontal flipping, random vertical flipping and random rotation; the strong enhancement strategy adds HSV shift on the basis of the weak enhancement strategy.

[0031] Furthermore, the calculated H&E matrix is ​​used to align the staining distributions in the feature space using an aH&E discriminator with a gradient reversal layer (GRL), including:

[0032] Improving the obtained discriminator to obtain an improved discriminator for predicting H&E components;

[0033] Change the binary cross entropy adversarial loss to:

[0034]

[0035] Among them, L adv represents the adversarial loss, L MSE Represents the square of the loss function L2; Represents the image in the strongly enhanced training dataset S H&E matrix; Represents the image in the strongly enhanced test dataset T H&E matrix;

[0036] Based on the improved discriminator, binary cross entropy adversarial loss is applied, and the gradient reversal layer (GRL) between the classification networks is combined to achieve adversarial learning.

[0037] Furthermore, the student model updates the teacher model through an exponential moving average, including:

[0038] The student model is updated by the back-propagation algorithm of gradient descent, and the weights of the teacher model are refined by the exponential moving average (EMA), specifically:

[0039]

[0040] In the above formula, represents the teacher model, t represents the number of iterations, and α represents the smoothing constant.

[0041] A method for constructing a deep learning system for semantic segmentation of colorectal adenoma recognition, comprising:

[0042] Acquire and process case images of colorectal adenoma to obtain the target data set;

[0043] Based on the dye called hematoxylin and eosin (H&E), combined with a semi-supervised learning model (Adversarial MeanTeacher, AMT), a semantic segmentation deep learning model for obtaining H&E Adversarial Mean Teacher (HEAMT) was constructed;

[0044] Based on the target dataset, the semantic segmentation deep model is trained.

[0045] Compared with the existing technology, the present invention has the following advantages and beneficial effects: by acquiring and processing case images of colorectal adenomas, and based on a dye called hematoxylin and eosin (H&E), combined with a semi-supervised learning model, a semantic segmentation deep learning model of H&E Adversarial Average Teacher (HEAMT) is constructed, and the semantic segmentation deep model is trained, which can ensure the creation quality of the semantic segmentation deep learning system for colorectal adenoma identification, and provide a basis for efficiently and accurately identifying case images of colorectal adenomas.

[0046] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings.

[0047] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0049] Figure 1 A schematic diagram of the structure of a semantic segmentation deep learning system for colorectal adenoma identification;

[0050] Figure 2 This is a schematic diagram of the data set acquisition and processing module structure;

[0051] Figure 3 A schematic diagram of the steps of a construction method. DETAILED DESCRIPTION

[0052] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0053] The present invention provides a colorectal adenoma recognition semantic segmentation deep learning system, such as Figure 1 Shown, including:

[0054] The data set acquisition and processing module is used to acquire and process case images of colorectal adenoma to obtain the target data set;

[0055] A model building module for building a semantic segmentation deep learning model based on a dye called hematoxylin and eosin (H&E) combined with a semi-supervised learning model (Adversarial Mean Teacher, AMT);

[0056] The model learning and training module is used to learn and train the semantic segmentation deep model based on the target dataset.

[0057] The working principle of the above technical solution is: in order to realize the construction of a semantic segmentation deep learning system for colorectal adenoma identification, the present invention proposes a data set acquisition and processing module for acquiring and processing case images of colorectal adenomas to obtain a target data set; a model construction module is proposed for constructing a semantic segmentation deep learning model of H&E adversarial mean teacher (HEAMT) based on a dye called hematoxylin and eosin (H&E) combined with a semi-supervised learning model (Adversarial Mean Teacher, AMT); a model learning and training module is proposed for learning and training the semantic segmentation deep model based on the target data set.

[0058] The beneficial effects of the above technical solution are: by adopting the solution provided in this embodiment, by acquiring and processing case images of colorectal adenomas, and based on a dye called hematoxylin and eosin (H&E), combined with a semi-supervised learning model, a semantic segmentation deep learning model of H&E Adversarial Average Teacher (HEAMT) is constructed, and the semantic segmentation deep model is trained, the creation quality of the semantic segmentation deep learning system for colorectal adenoma recognition can be guaranteed, providing a basis for efficiently and accurately identifying case images of colorectal adenomas.

[0059] In one embodiment, Figure 2 As shown, the data set acquisition and processing module includes a data set acquisition unit and a data set processing unit;

[0060] A data set acquisition unit is used to acquire case images of colorectal adenoma to form an original data set;

[0061] The data set processing unit is used to process the original data set to obtain the target data set.

[0062] The working principle of the above technical solution is: in order to achieve the acquisition and processing of the data set, the present invention first obtains case images of colorectal adenoma to form an original data set; then processes the original data set to obtain the target data set.

[0063] The beneficial effect of the above technical solution is: by adopting the solution provided by this embodiment, by acquiring and processing the original data set, a data basis is provided for the subsequent construction of the learning model.

[0064] In one embodiment, processing the original data set to obtain the target data set includes:

[0065] Divide the original data set into training data set and test data set;

[0066] Processing the training data set to obtain a processed training data set; wherein the processing includes color normalization processing and color enhancement using the HSV perturbation method; the color normalization processing is performed to make it have a color distribution similar to that of the test data set; the color enhancement is performed to expand the training data set;

[0067] Aggregate the test dataset and the processed training dataset to form the target dataset.

[0068] The working principle of the above technical solution is as follows: in order to realize the processing of the original data set, the present invention first divides the original data set to obtain a training data set and a test data set; then processes the training data set to obtain a processed training data set; wherein, the processing includes color normalization processing and color enhancement using the HSV perturbation method; color normalization processing is performed to make it have a color distribution similar to that of the test data set; color enhancement is performed to achieve the expansion of the training data set; finally, the test data set and the processed training data set are aggregated to form a target data set.

[0069] The beneficial effect of the above technical solution is: by adopting the solution provided by this embodiment, the availability of data processing can be guaranteed by processing the original data set, so as to enhance the learning and training effect of the learning model in learning and training.

[0070] In one embodiment, a semantic segmentation deep learning model for obtaining an H&E Adversarial Mean Teacher (HEAMT) is constructed based on a stain called hematoxylin and eosin (H&E) and combined with a semi-supervised learning model (Adversarial Mean Teacher, AMT), including:

[0071] Get the teacher (Mean Teacher, MT) model and the student model;

[0072] Combine the student model and the teacher model to obtain a deep learning model; the teacher model is used to generate high-quality pseudo labels to supervise the student model's learning in the target domain;

[0073] In the deep learning model, a H&E discriminator with a gradient reversal layer (GRL) is introduced to perform H&E-oriented adversarial learning, generating a semantic segmentation deep learning model with the H&E Adversarial Average Teacher (HEAMT). Among them, adversarial learning is used to guide the semantic segmentation deep learning model to learn coloring-invariant representation through a min-max game.

[0074] The working principle of the above technical solution is: in the process of constructing a semantic segmentation deep learning model for obtaining the H&E Adversarial Average Teacher (HEAMT), the present invention first obtains a teacher (Mean Teacher, MT) model and a student model; then the student model and the mean teacher model are combined to obtain a deep learning model; wherein, the teacher model is used to generate high-quality pseudo labels to supervise the student model's learning in the target domain; in the deep learning model, an aH&E discriminator with a gradient reversal layer (GRL) is introduced to perform H&E-oriented adversarial learning to generate a semantic segmentation deep learning model for obtaining the H&E Adversarial Average Teacher (HEAMT), wherein adversarial learning is used to guide the semantic segmentation deep learning model to learn coloring-invariant representation through a minimum-maximum game.

[0075] The beneficial effects of the above technical solution are: by adopting the solution provided in this embodiment, by obtaining a teacher model and a student model, and combining the student model and the average teacher model, a deep learning model is obtained, and finally an aH&E discriminator with a gradient reversal layer (GRL) is introduced to perform H&E-oriented adversarial learning, which can ensure the acquisition of a high-quality semantic segmentation deep learning model.

[0076] In one embodiment, learning and training a semantic segmentation deep model includes:

[0077] Set the training dataset S and test dataset T in the target dataset; where S includes N s Case images of labeled colorectal adenomas, in, Represents the i-th sample data in the training data set S; T includes N t Case images of unlabeled colorectal adenomas; in, Represents the i-th sample data in the training data set T;

[0078] According to the divided stages, based on the training dataset S and the test dataset T, the loss is optimized in an end-to-end manner to perform learning and training; the stages include the initialization stage and the mutual learning stage;

[0079] The initialization phase involves training a fully supervised classification model based on the training dataset S, i.e., training the classifier using labeled colorectal adenoma case images. The classifier is then replicated into two copies, which are used as the initialized student model and teacher model, respectively. The mutual learning phase involves the teacher model obtaining weakly enhanced images from the test dataset T, and the student model obtaining strongly enhanced images from the training dataset S and the test dataset T. The teacher model generates pseudo labels on the test dataset T to train the student model, which then updates the teacher model using an exponential moving average.

[0080] Using the calculated H&E matrix, an aH&E discriminator with a gradient reversal layer (GRL) is used to align the staining distributions in the feature space.

[0081] The working principle of the above technical solution is: in order to realize the learning and training of the semantic segmentation depth model, the present invention first sets the training data set S and the test data set T in the target data set; wherein S includes N s Case images of labeled colorectal adenomas, in, Represents the i-th sample data in the training data set S; T includes N t Case images of unlabeled colorectal adenomas; in, Represents the i-th sample data in the training dataset T; then according to the divided stages, based on the training dataset S and the test dataset T, the loss is optimized in an end-to-end manner to perform learning and training; the stages include an initialization stage and a mutual learning stage; among them, the initialization stage is: training a fully supervised classification model based on the training dataset S, that is, training the classifier with labeled case images of colorectal adenoma, and then copying it into two copies, and using the two copies as the initialized student model and teacher model respectively; the mutual learning stage is: the teacher model obtains weakly enhanced images from the test dataset T, and the student model obtains strongly enhanced images from the training dataset S and the test dataset T; the teacher model generates pseudo labels on the test dataset T to train the student model, and the student model updates the teacher model through the exponential moving average; finally, the calculated H&E matrix is ​​used to align the staining distribution in the feature space using an aH&E discriminator with a gradient reversal layer (GRL).

[0082] The beneficial effect of the above technical solution is: by adopting the solution provided by this embodiment, the quality of learning and training can be guaranteed by realizing learning and training of the semantic segmentation depth model.

[0083] In one embodiment, the mutual learning phase further includes:

[0084] Using strong and weak enhancement strategies, the student model is helped to learn coloring invariant features in the training dataset S and the test dataset T;

[0085] Inverse learning is introduced into the semantic segmentation deep learning model to align the distribution of the training dataset S and the test dataset T in the feature space.

[0086] The working principle of the above technical solution is: the mutual learning stage in the present invention also includes: using strong enhancement strategies and weak enhancement strategies to help the student model learn coloring invariant features in the training dataset S and the test dataset T; introducing inverse learning into the semantic segmentation deep learning model to align the distribution of the training dataset S and the test dataset T in the feature space.

[0087] The beneficial effect of the above technical solution is: by adopting the solution provided in this embodiment, by learning coloring invariant features and introducing inverse learning into the semantic segmentation deep learning model, the regularity of the feature distribution of the training dataset S and the test dataset T can be guaranteed.

[0088] In one embodiment, the weak enhancement strategy includes random horizontal flipping, random vertical flipping, and random rotation; the strong enhancement strategy adds HSV shift on the basis of the weak enhancement strategy.

[0089] The working principle of the above technical solution is: the weak enhancement strategy in the present invention includes random horizontal flipping, random vertical flipping and random rotation; the strong enhancement strategy adds HSV shift on the basis of the weak enhancement strategy.

[0090] The beneficial effect of the above technical solution is: by adopting the solution provided by this embodiment, through the combined use of weak enhancement strategy and strong enhancement strategy, it is possible to ensure that the coloring-invariant features are learned.

[0091] In one embodiment, the calculated H&E matrix is ​​used to align the staining distributions in the feature space using an aH&E discriminator with a gradient reversal layer (GRL), including:

[0092] Improving the obtained discriminator to obtain an improved discriminator for predicting H&E components;

[0093] Change the binary cross entropy adversarial loss to:

[0094]

[0095] Among them, L adv represents the adversarial loss, L MSE Represents the square of the loss function L2; Represents the image in the strongly enhanced training dataset S H&E matrix; Represents the image in the strongly enhanced test dataset T H&E matrix;

[0096] Based on the improved discriminator, binary cross entropy adversarial loss is applied, and the gradient reversal layer (GRL) between the classification networks is combined to achieve adversarial learning.

[0097] The working principle of the above technical solution is as follows: due to the lack of annotations on the target data, pseudo labels in the teacher model are used for self-training. However, since the supervision through the initialization stage only comes from the training dataset S, the initial model at the beginning of the mutual learning stage may have a strong source bias, and this domain bias needs to be compensated; traditional adversarial learning involves two important parts: a feature extractor and a domain discriminator; the domain discriminator is used to distinguish features from the training dataset S or the test dataset T and predict the probability of the test dataset T for each sample, and then the binary cross-entropy adversarial loss Ladv is applied to the probability-based label image; in order to realize the use of the calculated H&E matrix, a H&E discriminator with a gradient reversal layer (GRL) is used to align the staining distribution in the feature space. The present invention first improves the obtained discriminator to obtain an improved discriminator to predict the H&E component; then the binary cross-entropy adversarial loss is changed, and then, based on the improved discriminator, the binary cross-entropy adversarial loss is applied, combined with the gradient reversal layer (GRL) between the classification networks to achieve adversarial learning.

[0098] The beneficial effects of the above technical solution are: by adopting the solution provided by this embodiment, by improving the obtained discriminator and then changing the binary cross-entropy adversarial loss, it is possible to learn visual representations that are independent of coloring changes, thereby aligning images of different colors into a common space, and then, based on the improved discriminator, applying the binary cross-entropy adversarial loss, combined with the gradient reversal layer (GRL) between the classification networks to achieve adversarial learning, which can ensure the quality of adversarial learning.

[0099] In one embodiment, the student model updates the teacher model via an exponential moving average, including:

[0100] The student model is updated by the back-propagation algorithm of gradient descent, and the weights of the teacher model are refined by the exponential moving average (EMA), specifically:

[0101]

[0102] In the above formula, represents the teacher model, t represents the number of iterations, and α represents the smoothing constant.

[0103] The working principle of the above technical solution is: the student model in the present invention updates the teacher model through the exponential moving average, which is reflected in: the student model is updated by the back propagation algorithm of gradient descent, and the weight of the teacher model is refined by the exponential moving average (EMA).

[0104] The beneficial effect of the above technical solution is: using the solution provided by this embodiment, the teacher model can be regarded as a collection of different versions of learning models in the entire pre-training process, so the robustness and stability of the teacher model are guaranteed, and accurate false labels can be generated.

[0105] A construction method such as Figure 3 As shown in the figure, a deep learning system for semantic segmentation of colorectal adenoma recognition is constructed, including:

[0106] Acquire and process case images of colorectal adenoma to obtain the target data set;

[0107] Based on the dye called hematoxylin and eosin (H&E), combined with a semi-supervised learning model (Adversarial MeanTeacher, AMT), a semantic segmentation deep learning model for obtaining H&E Adversarial Mean Teacher (HEAMT) was constructed;

[0108] Based on the target dataset, the semantic segmentation deep model is trained.

[0109] The working principle of the above technical solution is as follows: in order to implement the construction method, the present invention first obtains and processes case images of colorectal adenomas to obtain a target data set; then, based on a dye called hematoxylin and eosin (H&E), combined with a semi-supervised learning model (Adversarial Mean Teacher, AMT), a semantic segmentation deep learning model of H&E Adversarial Mean Teacher (HEAMT) is constructed; finally, based on the target data set, the semantic segmentation deep model is learned and trained.

[0110] The beneficial effects of the above technical solution are: by adopting the solution provided in this embodiment, by acquiring and processing case images of colorectal adenomas, and based on a dye called hematoxylin and eosin (H&E), combined with a semi-supervised learning model, a semantic segmentation deep learning model of H&E Adversarial Average Teacher (HEAMT) is constructed, and the semantic segmentation deep model is trained, the creation quality of the semantic segmentation deep learning system for colorectal adenoma recognition can be guaranteed, providing a basis for efficiently and accurately identifying case images of colorectal adenomas.

[0111] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A deep learning system for colorectal adenoma recognition and semantic segmentation, characterized by: include: The data set acquisition and processing module is used to acquire and process case images of colorectal adenoma to obtain the target data set; A model building module for building a semantic segmentation deep learning model based on a dye called hematoxylin and eosin (H&E) combined with a semi-supervised learning model (Adversarial Mean Teacher, AMT); The model learning and training module is used to learn and train the semantic segmentation deep model based on the target dataset; it includes: Set the training dataset S and test dataset T in the target dataset; where S includes N s Case images of labeled colorectal adenomas, in, Represents the i-th sample data in the training data set S; T includes N t Case images of unlabeled colorectal adenomas; in, Represents the i-th sample data in the training data set T; According to the divided stages, based on the training dataset S and the test dataset T, the loss is optimized in an end-to-end manner to perform learning and training; the stages include the initialization stage and the mutual learning stage; The initialization phase involves training a fully supervised classification model based on the training dataset S, i.e., training the classifier using labeled colorectal adenoma case images. The classifier is then replicated into two copies, which are used as the initialized student model and teacher model, respectively. The mutual learning phase involves the teacher model obtaining weakly enhanced images from the test dataset T, and the student model obtaining strongly enhanced images from the training dataset S and the test dataset T. The teacher model generates pseudo labels on the test dataset T to train the student model, which then updates the teacher model using an exponential moving average. The mutual learning phase also includes: using strong and weak augmentation strategies to help the student model learn color-invariant features in the training dataset S and the test dataset T; the weak augmentation strategy includes random horizontal flip, random vertical flip, and random rotation; the strong augmentation strategy adds HSV shift on the basis of the weak augmentation strategy; Inverse learning is introduced into the semantic segmentation deep learning model to align the distribution of the training dataset S and the test dataset T in the feature space; Using the calculated H&E matrix, an aH&E discriminator with a gradient reversal layer (GRL) is used to align the staining distributions in the feature space; including: Improving the obtained discriminator to obtain an improved discriminator for predicting H&E components; Change the binary cross entropy adversarial loss to: Among them, L adv represents the adversarial loss, L MSE Represents the square of the loss function L2; Represents the image in the strongly enhanced training dataset S H&E matrix; Represents the image in the strongly enhanced test dataset T H&E matrix; Based on the improved discriminator, binary cross entropy adversarial loss is applied, and the gradient reversal layer (GRL) between the classification networks is combined to achieve adversarial learning.

2. A colorectal adenoma recognition semantic segmentation deep learning system according to claim 1, characterized in that: The data set acquisition and processing module includes a data set acquisition unit and a data set processing unit; A data set acquisition unit is used to acquire case images of colorectal adenoma to form an original data set; The data set processing unit is used to process the original data set to obtain the target data set.

3. A colorectal adenoma recognition semantic segmentation deep learning system according to claim 2, characterized in that: Process the original data set to obtain the target data set, including: Divide the original data set into training data set and test data set; Processing the training data set to obtain a processed training data set; wherein the processing includes color normalization processing and color enhancement using the HSV perturbation method; the color normalization processing is performed to make it have a color distribution similar to that of the test data set; the color enhancement is performed to expand the training data set; Aggregate the test dataset and the processed training dataset to form the target dataset.

4. A colorectal adenoma recognition semantic segmentation deep learning system according to claim 1, characterized in that: Based on the dye called hematoxylin and eosin (H&E), combined with a semi-supervised learning model (Adversarial Mean Teacher, AMT), a semantic segmentation deep learning model for obtaining H&E Adversarial Mean Teacher (HEAMT) was constructed, including: Get the teacher (Mean Teacher, MT) model and the student model; Combine the student model and the teacher model to obtain a deep learning model; the teacher model is used to generate high-quality pseudo labels to supervise the student model's learning in the target domain; In the deep learning model, a H&E discriminator with a gradient reversal layer (GRL) is introduced to perform H&E-oriented adversarial learning, generating a semantic segmentation deep learning model with the H&E Adversarial Average Teacher (HEAMT). Among them, adversarial learning is used to guide the semantic segmentation deep learning model to learn coloring-invariant representation through a min-max game.

5. The colorectal adenoma recognition semantic segmentation deep learning system according to claim 1, characterized in that: The student model updates the teacher model through an exponential moving average, including: The student model is updated by the back-propagation algorithm of gradient descent, and the weights of the teacher model are refined by the exponential moving average (EMA), specifically: In the above formula, represents the teacher model, t represents the number of iterations, and α represents the smoothing constant.

6. A construction method, characterized in that: Used to construct a colorectal adenoma recognition semantic segmentation deep learning system according to any one of claims 1 to 5, characterized by comprising: Acquire and process case images of colorectal adenoma to obtain the target data set; Based on the dye called hematoxylin and eosin (H&E), combined with a semi-supervised learning model (Adversarial MeanTeacher, AMT), a semantic segmentation deep learning model for obtaining H&E Adversarial Mean Teacher (HEAMT) was constructed; Based on the target dataset, the semantic segmentation deep model is trained.

Images