Colonic polyp image segmentation method based on cross-domain entropy enhanced prototype contrast learning

Abstract

The application discloses a colon polyp image segmentation method based on cross-domain entropy enhanced prototype contrast learning, comprising the following steps: constructing a network model CEPCL, which is composed of a source guided cross-domain representation contrast learning module including a first encoder and a projection head, and an entropy weighted prototype contrast learning module including a segmentation network, an entropy filtering module and a class prototype extraction module; initializing the parameters of CEPCL and training; the first encoder extracts the features of the source domain and target domain data and maps them into source domain and target domain projection features through the projection head, constructs cross-domain positive / negative sample pairs with the source domain projection features as anchor points, and calculates the pre-training contrast loss; the second encoder shares the weights of the first encoder and extracts the features of the source domain and target domain data, the decoder outputs the source domain and target domain prediction probability map, and the entropy filtering module extracts the class prototype; an entropy weighted contrast loss optimization mechanism is introduced, the entropy weighted prototype contrast loss and the source domain supervised segmentation loss are calculated; the total loss is calculated; and the segmentation capability in the cross-domain scene is improved.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically a method for segmenting colon polyp images based on cross-domain entropy-enhanced prototype contrastive learning. Background Technology

[0002] Colorectal cancer is one of the malignant tumors with high incidence and mortality rates worldwide. Its high incidence is closely related to factors such as diet and genetics. Compared with other cancers, the prognosis of colorectal cancer is closely related to early screening. Early detection and treatment can significantly improve the cure rate, especially for high-risk groups. Regular screening can effectively reduce the mortality rate of colorectal cancer.

[0003] Polyps, as an early form of colorectal cancer, are effectively segmented for early screening. In recent years, fully supervised learning methods have made significant progress in medical image segmentation; however, they heavily rely on large-scale, high-quality pixel-level labeled data. In real-world clinical scenarios, the extremely high cost and long acquisition time for labeling colon polyps significantly limit the application of fully supervised learning. Semi-supervised segmentation methods, by jointly utilizing a small number of labeled samples with a large number of unlabeled samples, effectively reduce the reliance on manual annotation and have become a research hotspot in the field of medical image segmentation.

[0004] Currently, contrastive learning, by bringing similar samples closer together and pushing away dissimilar samples, improves the discriminative power of features and is widely used in semi-supervised medical image segmentation. For example, Xiao et al. proposed the contrastive learning-driven framework CTNet, which introduces supervised pixel contrastive learning into the Transformer backbone network to enhance intra-class aggregation and inter-class separation in the feature space; Sebai proposed DiffusionMixNet, which uses a base model to assist in locating uncertain regions and corrects edge features through local intra-image contrastive learning; in addition, Li et al. proposed RD-Net, which enhances the network's complementary representation ability in difficult-to-segment regions through deep-guided cross-modal contrastive mutual learning.

[0005] While the aforementioned semi-supervised methods have demonstrated good segmentation performance in specific scenarios, they still have two shortcomings in real cross-domain scenarios: First, there are significant differences in imaging conditions and distribution shifts between the source and target domains. Existing contrastive learning strategies often struggle to achieve effective pixel-level cross-domain feature alignment without target domain labels, leading to a degradation in the model's generalization ability in unseen domains. Second, when constructing contrastive samples, they are easily affected by pseudo-label noise from the target domain and lack an effective screening mechanism for reliable samples. This makes it difficult for the model to ensure the stability of class structure discrimination when facing polyps with low contrast, blurred boundaries, or varied morphologies. Summary of the Invention

[0006] The purpose of this invention is to provide a colon polyp image segmentation method based on cross-domain entropy-enhanced prototype contrastive learning, which has excellent colon polyp segmentation accuracy and cross-domain generalization ability under minimal annotation conditions.

[0007] This invention is achieved through the following technical solution: A method for segmenting colon polyps based on cross-domain entropy-enhanced prototype contrastive learning includes the following steps: Step 1: Construct the medical image segmentation network model CEPCL, which includes a source-guided cross-domain representation contrastive learning module and an entropy-weighted prototype contrastive learning module; The source-guided cross-domain representation contrastive learning module includes a first encoder and a projection head, wherein: the first encoder is used to extract source domain labeled data. and unlabeled data in the target domain Pixel-level features are used to generate source domain pixel-level feature maps. and target domain pixel-level feature map The projection head is used to project pixel-level feature maps from the source domain. and target domain pixel-level feature map Each pixel feature in the image is mapped to a high-dimensional feature space to generate source domain projection features. and target domain projection features Based on source domain projection features Using anchor points, project features into the target domain. In the selection process, features of the same category and pixel features of different categories are used as positive samples respectively. and negative samples This allows for the construction of cross-domain positive and negative sample pairs, respectively. The entropy-weighted prototype comparison learning module includes a segmentation network, an entropy filtering module, and a class prototype extraction module, wherein: the segmentation network includes a second encoder and two decoders, the second encoder being connected to the first encoder via weight sharing, and the second encoder being used to extract source domain labeled data. and unlabeled data in the target domain Deep pixel-level features are used to generate source domain pixel-level feature maps. and target domain pixel-level feature map Two decoders are used to decode the source domain pixel-level feature maps respectively. and target domain pixel-level feature map Output the source domain prediction probability map Target Domain Prediction Probability Map The entropy filtering module is based on the source domain prediction probability map. Target Domain Prediction Probability Map The probability distribution in the graph is used to calculate the prediction entropy of each pixel, which measures pixel uncertainty. A set of reliable pixels with high confidence is then selected through a preset entropy threshold. The class prototype extraction module extracts data from a reliable pixel set. The middle belongs to the first Extracting high-quality class prototypes from pixel features of a class ; Step 2: Initialize the parameters of the medical image segmentation network model CEPCL and train it; Step 3: In the source-guided cross-domain representation contrastive learning stage, the source domain labeled data... and unlabeled data in the target domain The first encoder extracts pixel-level features, which are then mapped by the projection head to generate source domain projection features. and target domain projection features In high-dimensional space, the source domain projection features Construct cross-domain positive and negative sample pairs for anchor points and calculate the pre-trained contrastive loss. ; Step 4: In the entropy-weighted prototype contrastive learning stage, based on the cross-domain representation pre-training results completed in Step 3, the stable feature extraction parameters of the first encoder are transferred to the second encoder through a weight sharing mechanism, achieving efficient transfer of pre-trained features to the segmentation task. Source domain labeled data... Unlabeled data in the target domain The input segmentation network is synchronously fed into a second encoder, which extracts deep pixel-level features that combine domain invariance and task specificity. These features are then independently decoded by dual decoders, outputting corresponding prediction probability maps. An entropy filtering module quantifies the uncertainty and filters reliable regions within the prediction probability maps, removing low-confidence and high-noise interference pixels to obtain a pixel set containing only high-reliability predictions. Based on this reliable pixel set, features are aggregated and calculated by category to generate high-quality category prototypes that represent the core semantics of each category, providing stable category center constraints for subsequent pixel-prototype comparison learning. Step 5: Using category prototypes To guide this process, an entropy-weighted contrastive loss optimization mechanism is introduced to calculate the entropy-weighted prototype contrastive loss. and source domain supervision segmentation loss ; Step 6: Calculate the total loss , is represented as: + In the formula: These are weighting coefficients; , and These are source domain supervised segmentation loss, pre-training contrastive loss, and class prototype contrastive loss, respectively.

[0008] Furthermore, the process of initializing the parameters of the medical image segmentation network model CEPCL and training it in step 2 includes: using a stochastic gradient descent (SGD) optimizer to train CEPCL and update the parameters, setting the weight decay factor to 0.0005, the momentum to 0.9, the initial learning rate to 0.01, using a multinomial scheduling strategy during training to reduce the learning rate to 0.001, setting the input image size to 224×224, and the batch size to 8.

[0009] Furthermore, the pre-trained contrastive loss in step 3... , is represented as: In the formula: Represents the projection features of the target domain Pixel features that are of the same category as the anchor point are considered positive samples. Projection features of the target domain Pixel features of different categories from the anchor point are considered negative samples. Indicates anchor point Compared with positive samples The similarity between them; Indicates anchor point With negative samples The similarity between them; It is the temperature coefficient.

[0010] Furthermore, the data processing procedure of the entropy filtering module in step 4 is as follows: Step 4.1: Define each pixel Predicted entropy , is represented as: In the formula: For pixels Belongs to the The predicted probability of a class; Total number of categories; Step 4.2: Pass the preset entropy threshold. Filtering a reliable set of pixels with high confidence , is represented as: In the formula: Entropy threshold; A reliable set of pixels; Step 4.3: Utilize a reliable pixel set The Middle Calculation of the mean pixel feature of the class High-quality class prototypes , is represented as: In the formula: For the first A reliable set of pixels; For pixels eigenvectors.

[0011] Furthermore, step 5 is as follows: Step 5.1, using pixel-level features As an anchor point, construct the category prototype. Guided pixel-level contrastive learning loss is applied and entropy-weighted optimization is performed. The specific process is as follows: Step 5.1.1: Using pixel-level features As anchor points, prototypes of the same category as the target are used as positive samples, and the rest of the prototype sets are used as positive samples. As negative samples, pixel-level comparison terms are constructed. , is represented as: In the formula: For pixels eigenvectors; For pixels The prediction category; For the first The class's category prototype; For pixels The predicted category corresponds to the category prototype; It is the temperature coefficient; express and The similarity between them; for and The similarity between them; Step 5.1.2: Introduce an entropy-weighted strategy and define pixel weights. , is represented as: In the formula: For pixels Entropy weights; For pixels Predicted entropy; Represents pixels Predicted entropy; Represents any pixel in the current batch; Represents the number of pixels in the current batch. The maximum value of the predicted entropy; Step 5.1.3: Calculate the entropy-weighted class-prototype contrast loss. , is represented as: In the formula: This represents the set of pixels that participate in the contrastive learning process. Pixel-level contrast; Pixel weights; Step 5.2: Calculate the source domain supervised segmentation loss. , is represented as: In the formula: Provide the actual segmentation label for the source domain; Represents cross-entropy loss; Indicates Dice loss; for Weighting coefficients; for Weighting coefficients; This is the source domain prediction probability map.

[0012] Furthermore, in step 6 The value of is 0.1.

[0013] The present invention has the following beneficial technical effects: This invention can accurately determine the contour position of the target to be segmented in colon images and accurately segment the image. This is reflected in the following aspects: First, a source-guided cross-domain representation contrastive learning module is designed, using source domain category information as anchor points to guide the effective alignment of target domain pixel features in the feature space, reducing feature bias caused by distribution offset and enhancing the network's adaptability to cross-domain scenarios. Second, an entropy-weighted prototype contrastive learning module uses the entropy of the predicted distribution to measure pixel uncertainty and selects reliable pixels to construct quality category prototypes. Through contrastive learning between pixels and class prototypes, intra-class aggregation and inter-class separation are promoted, enhancing the network's fine-grained discrimination ability and boundary stability of the target domain. Third, an entropy-weighted contrastive loss optimization mechanism is constructed. By assigning greater weight to high-entropy uncertain regions, targeted optimization is achieved for difficult samples and ambiguous boundary regions, enabling accurate extraction of the contour position of the target to be segmented in medical images and improving the network's robustness and segmentation accuracy under complex distributions. In short, this invention effectively alleviates the distribution shift problem of the network model in cross-domain scenarios through a two-stage progressive training of the source-guided cross-domain representation contrastive learning module and the entropy-weighted prototype contrastive learning module, and significantly improves the model's colon polyp segmentation accuracy and cross-domain generalization ability under conditions with very few annotations. Attached Figure Description

[0014] Figure 1 : A schematic diagram of the structure of the CEPCL network of the present invention; Figure 2 Visual comparison of the results obtained by segmenting a public polyp dataset using the CEPCL network of this invention and existing networks. Detailed Implementation

[0015] The present invention will be further described in detail below with reference to specific embodiments. These descriptions are for explanation purposes only and are not intended to limit the scope of the invention.

[0016] A method for segmenting colon polyps based on cross-domain entropy-enhanced prototype contrastive learning includes the following steps: Step 1: Introduce a source-guided cross-domain representation contrastive learning strategy and an entropy-enhanced prototype contrastive learning strategy into the network containing an encoder, decoder, projector head, and class prototype extraction module, thereby constructing a network such as... Figure 1 The medical image segmentation network model CEPCL shown includes a source-guided cross-domain representation contrastive learning module and an entropy-weighted prototype contrastive learning module. The source-guided cross-domain representation contrastive learning module includes a first encoder and a projection head, wherein: the first encoder is used to extract source domain labeled data. and unlabeled data in the target domain Pixel-level features are used to generate source domain pixel-level feature maps. and target domain pixel-level feature map The projection head is used to project pixel-level feature maps from the source domain. and target domain pixel-level feature map Each pixel feature in the image is mapped to a high-dimensional feature space to generate source domain projection features. and target domain projection features Based on source domain projection features Using anchor points, project features into the target domain. In the selection process, features of the same category and pixel features of different categories are used as positive samples respectively. and negative samples This allows for the construction of cross-domain positive and negative sample pairs, respectively, and the calculation of the pre-training contrastive loss of the source-guided cross-domain representation contrastive learning module. To achieve pixel-level cross-domain distribution alignment; The entropy-weighted prototype comparison learning module includes a segmentation network, an entropy filtering module, and a class prototype extraction module. The segmentation network includes a second encoder and two decoders. The second encoder is connected to the first encoder via weight sharing, meaning the second encoder inherits the pre-trained weights of the first encoder. The second encoder is used to extract source domain labeled data. and unlabeled data in the target domain Deep pixel-level features are used to generate source domain pixel-level feature maps. and target domain pixel-level feature map Two decoders are used to decode the source domain pixel-level feature maps respectively. and target domain pixel-level feature map It also outputs the source domain prediction probability map. Target Domain Prediction Probability Map The entropy filtering module is based on the source domain prediction probability map. Target Domain Prediction Probability Map The probability distribution in the graph is used to calculate the prediction entropy of each pixel, which measures pixel uncertainty. A set of reliable pixels with high confidence is then selected through a preset entropy threshold. The class prototype extraction module extracts data from a reliable pixel set. The middle belongs to the first Extracting high-quality class prototypes from pixel features of a class This leads to the construction of a pixel-level class prototype contrastive learning loss. In order to give greater weight to pixels in uncertain regions, an entropy weighting strategy is introduced to achieve targeted optimization for difficult samples and ambiguous boundary regions. Step 2: Initialize the parameters of the medical image segmentation network model CEPCL and train it. The process includes: using the stochastic gradient descent (SGD) optimizer to train CEPCL and update the parameters, where: the weight decay factor is set to 0.0005, the momentum is set to 0.9, the initial learning rate is set to 0.01, a multinomial scheduling strategy is used during training to reduce the learning rate to 0.001, the input image size is set to 224×224, and the batch size is set to 8. Step 3: In the source-guided cross-domain representation contrastive learning stage, the source domain labeled data... and unlabeled data in the target domain The first encoder extracts pixel-level features, which are then mapped by the projection head to generate source domain projection features. and target domain projection features In high-dimensional space, the source domain projection features Construct cross-domain positive and negative sample pairs for anchor points and calculate the pre-trained contrastive loss. , is represented as: In the formula: Represents the projection features of the target domain Pixel features that are of the same category as the anchor point are considered positive samples. Projection features of the target domain Pixel features of different categories from the anchor point are considered negative samples. Indicates anchor point Compared with positive samples The similarity between them; Indicates anchor point With negative samples The similarity between them; It is the temperature coefficient; Step 4: In the entropy-weighted prototype contrastive learning stage, based on the cross-domain representation pre-training results completed in Step 3, the stable feature extraction parameters of the first encoder are transferred to the second encoder through a weight sharing mechanism, achieving efficient transfer of pre-trained features to the segmentation task. Source domain labeled data... Unlabeled data in the target domain The input segmentation network is synchronously fed into a second encoder, which extracts deep pixel-level features that combine domain invariance and task specificity. These features are then independently decoded by dual decoders, outputting corresponding prediction probability maps. An entropy filtering module quantifies the uncertainty and filters reliable regions within the prediction probability maps, removing low-confidence and high-noise interference pixels to obtain a pixel set containing only high-reliability predictions. Based on this reliable pixel set, features are aggregated and calculated by category to generate high-quality category prototypes that represent the core semantics of each category, providing stable category center constraints for subsequent pixel-prototype comparison learning. The data processing procedure of the entropy filtering module is as follows: Step 4.1: Define each pixel Predicted entropy , is represented as: In the formula: For pixels Belongs to the The predicted probability of a class; Total number of categories; Step 4.2: Pass the preset entropy threshold. Filtering a reliable set of pixels with high confidence , is represented as: In the formula: Entropy threshold; A reliable set of pixels; Step 4.3: Utilize a reliable pixel set The Middle Calculation of the mean pixel feature of the class High-quality class prototypes , is represented as: In the formula: For the first A reliable set of pixels; For pixels eigenvectors; Step 5: Using category prototypes To guide this process, an entropy-weighted contrastive loss optimization mechanism is introduced to calculate the entropy-weighted prototype contrastive loss. and source domain supervision segmentation loss The process is as follows: Step 5.1, using pixel-level features As an anchor point, a pixel-level contrastive learning loss guided by category prototypes is constructed and optimized using entropy weighting. The specific process is as follows: Step 5.1.1: Using pixel-level features As anchor points, prototypes of the same category as the target are used as positive samples, and the rest of the prototype sets are used as positive samples. As negative samples, pixel-level comparison terms are constructed. , is represented as: In the formula: For pixels The feature vector, i.e., pixel-level features; For pixels The prediction category; For the first The class's category prototype; For pixels The predicted category corresponds to the category prototype, i.e., the pixel. Which category it is predicted to belong to, then This represents the category prototype corresponding to that category; It is the temperature coefficient; express and The similarity between them; for and The similarity between them; Step 5.1.2: Introduce an entropy-weighted strategy and define pixel weights. , is represented as: In the formula: For pixels Entropy weights; For pixels Predicted entropy; Represents pixels Predicted entropy; Represents any pixel in the current batch; Represents the number of pixels in the current batch. The maximum value of the predicted entropy is used for normalization; Step 5.1.3: Calculate the entropy-weighted class-prototype contrast loss. , is represented as: In the formula: This represents the set of pixels that participate in the contrastive learning process. Pixel-level contrast; Pixel weights; Step 5.2: Calculate the source domain supervised segmentation loss. , is represented as: In the formula: Provide the actual segmentation label for the source domain; Represents cross-entropy loss; Indicates Dice loss; for Weighting coefficients; for Weighting coefficients; Predict the probability map for the source domain; Step 6: Calculate the total loss , is represented as: + In the formula: These are weighting coefficients, all set to 0.1; Source domain supervision segmentation loss; For pre-training contrastive loss; The loss is a prototype-based comparison loss.

[0017] The medical image segmentation network model CEPCL proposed in this embodiment uses the PyTorch deep learning framework. To verify the performance of the proposed medical image segmentation network model CEPCL on polyp segmentation, the network was trained and tested using the publicly available datasets Kvasir-SEG, CVC-ClinicDB, CVC-300, CVC-ColonDB, and ETIS-LaribPolypDB. The training set consisted of Kvasir-SEG and CVC-ClinicDB, serving as the source domain data for model learning. The test set consisted of CVC-300, CVC-ColonDB, and ETIS-LaribPolypDB, serving as the target domain data for model generalization validation. These training and test sets originated from different clinical acquisition devices, imaging environments, and patient groups, exhibiting inter-domain differences in imaging style, resolution, noise distribution, and lesion appearance, representing typical cross-domain distributed data. The cross-domain testing referred to in this embodiment involves performing inference evaluation directly on an unseen target domain dataset with a different distribution after model training on the source domain dataset. This is used to test the model's cross-domain generalization ability and segmentation robustness under conditions of no target domain annotation and domain offset.

[0018] The verification environment was as follows: CPU: Intel(R) Xeon(R) Gold 6226R; RAM: 32 GB; GPU: Nvidia GeForce RTX 3090 with 24 GB of video memory, and it was conducted on the Ubuntu 16.04.10 operating system.

[0019] The segmentation performance is evaluated using the following metrics: In the formula: , and represent the number of true negatives, true negatives, and false negatives, respectively. and It is typically used to calculate the similarity between two regions; The value range is [0, 1]. The closer the value is to 1, the better the segmentation effect.

[0020] Ablation experiments were used to verify the source-guided cross-domain representation contrastive learning strategy in the proposed medical image segmentation network model CEPCL in this embodiment. Comparative learning strategies with entropy-weighted prototypes The role of this is to use U-Net, which is trained under supervision using labeled data, as the baseline model (see the first row of Table 1 for details), and gradually introduce... and The effectiveness of the source-guided cross-domain representation contrastive pre-training module and the entropy-weighted prototype contrastive learning module was validated on the CVC-300 dataset. The ablation experiment results are shown in Table 1, where: This represents the source domain supervised segmentation loss of the baseline model; This indicates a source-guided cross-domain representation contrastive learning module; This indicates the use of the entropy-weighted prototype contrastive learning module; the first row in Table 1 shows the ablation test results of the baseline model under different annotation ratios.

[0021] As can be seen from Table 1: Introducing... and Then, the model was annotated at three different percentages: 5%, 12.5%, and 25%. mDice and mIoU The metrics show significant improvements compared to the baseline model. This is attributed to the source domain-guided cross-domain contrast constraint, which effectively mitigates feature drift caused by distribution shift, while prototype contrast learning enhances intra-class feature aggregation and inter-class separation. When the two strategies are applied together, the model achieves optimal values ​​on all metrics, demonstrating a significant synergistic gain effect. This illustrates the effectiveness of the proposed CEPCL model in cross-domain medical image segmentation tasks.

[0022] Table 1 Ablation Experiment Results Table 2 shows a comparison of the segmentation results of the proposed medical image segmentation network model CEPCL with mainstream semi-supervised segmentation network models in recent years on existing public datasets for polyp images. Bold text represents the optimal results. As shown in Table 2, CEPCL achieves the best results on five polyp datasets: Kvasir-SEG, CVC-ClinicDB, CVC-300, CVC-ColonDB, and ETIS-LaribPolypDB. Especially on the challenging target domain datasets CVC-ColonDB and ETIS, CEPCL demonstrates a significant segmentation advantage over the suboptimal network DFCPS. mDice The values ​​are 5.13% and 2.43% higher, respectively; on the Kvasir-SEG and CVC-300 datasets, they are significantly better than the suboptimal network's DFCPS. mDice The values ​​are 0.86% and 3.05% higher; it also outperforms the suboptimal PedSemiSeg network on the CVC-ClinicDB dataset. mDice The value is 0.59% higher. In summary, compared with existing advanced networks, CEPCL has stronger segmentation accuracy and cross-domain generalization, and can accurately segment polyp boundaries.

[0023] Table 2. Segmentation results of polyp images from existing public datasets using different network models. Figure 2 This document presents visualizations of the proposed network model CEPCL and existing network models on a polyp segmentation dataset. Figure 2 It can be seen that CEPCL has robust segmentation capabilities when facing polyps of different shapes and sizes in cross-domain scenarios, and its cross-domain adaptability and segmentation accuracy are better than other models. In addition, CEPCL has a superior perception capability for blurred edges and low-contrast regions of polyps through an entropy weighting mechanism, and can extract the boundary contours of polyp targets more accurately.

Claims

1. A colon polyp image segmentation method based on cross-domain entropy-enhanced prototype contrastive learning, characterized in that, Includes the following steps: Step 1: Construct the medical image segmentation network model CEPCL, which includes a source-guided cross-domain representation contrastive learning module and an entropy-weighted prototype contrastive learning module; The source-guided cross-domain representation contrastive learning module includes a first encoder and a projection head, wherein: the first encoder is used to extract source domain labeled data. and unlabeled data in the target domain Pixel-level features are used to generate source domain pixel-level feature maps. and target domain pixel-level feature map The projection head is used to project pixel-level feature maps from the source domain. and target domain pixel-level feature map Each pixel feature in the image is mapped to a high-dimensional feature space to generate source domain projection features. and target domain projection features Based on source domain projection features Using anchor points, project features into the target domain. In the selection process, features of the same category and pixel features of different categories are used as positive samples respectively. and negative samples This allows for the construction of cross-domain positive and negative sample pairs, respectively. The entropy-weighted prototype comparison learning module includes a segmentation network, an entropy filtering module, and a class prototype extraction module, wherein: the segmentation network includes a second encoder and two decoders, the second encoder being connected to the first encoder via weight sharing, and the second encoder being used to extract source domain labeled data. and unlabeled data in the target domain Deep pixel-level features are used to generate source domain pixel-level feature maps. and target domain pixel-level feature map Two decoders are used to decode the source domain pixel-level feature maps respectively. and target domain pixel-level feature map Output the source domain prediction probability map Target Domain Prediction Probability Map The entropy filtering module is based on the source domain prediction probability map. Target Domain Prediction Probability Map The probability distribution in the graph is used to calculate the prediction entropy of each pixel, which measures pixel uncertainty. A set of reliable pixels with high confidence is then selected through a preset entropy threshold. The class prototype extraction module extracts data from a reliable pixel set. The middle belongs to the first Extracting high-quality class prototypes from pixel features of a class ; Step 2: Initialize the parameters of the medical image segmentation network model CEPCL and train it; Step 3: In the source-guided cross-domain representation contrastive learning stage, the source domain labeled data... and unlabeled data in the target domain The first encoder extracts pixel-level features, which are then mapped by the projection head to generate source domain projection features. and target domain projection features In high-dimensional space, the source domain projection features Construct cross-domain positive and negative sample pairs for anchor points and calculate the pre-trained contrastive loss. ; Step 4: In the entropy-weighted prototype contrastive learning stage, based on the cross-domain representation pre-training results completed in Step 3, the stable feature extraction parameters of the first encoder are transferred to the second encoder through a weight sharing mechanism, achieving efficient transfer of pre-trained features to the segmentation task. Source domain labeled data... Unlabeled data in the target domain The input segmentation network is synchronously fed into a second encoder, which extracts deep pixel-level features that combine domain invariance and task specificity. These features are then independently decoded by dual decoders, outputting corresponding prediction probability maps. An entropy filtering module quantifies the uncertainty and filters reliable regions within the prediction probability maps, removing low-confidence and high-noise interference pixels to obtain a pixel set containing only high-reliability predictions. Based on this reliable pixel set, features are aggregated and calculated by category to generate high-quality category prototypes that represent the core semantics of each category, providing stable category center constraints for subsequent pixel-prototype comparison learning. Step 5: Using category prototypes To guide this process, an entropy-weighted contrastive loss optimization mechanism is introduced to calculate the entropy-weighted prototype contrastive loss. and source domain supervision segmentation loss ; Step 6: Calculate the total loss , represented as: + In the formula: These are weighting coefficients; , and These are source domain supervised segmentation loss, pre-training contrastive loss, and class prototype contrastive loss, respectively.

2. The colon polyp image segmentation method based on cross-domain entropy-enhanced prototype contrastive learning according to claim 1, characterized in that, The process of initializing the parameters of the medical image segmentation network model CEPCL and training it in step 2 includes: using the stochastic gradient descent (SGD) optimizer to train CEPCL and update the parameters, setting the weight decay factor to 0.0005, the momentum to 0.9, the initial learning rate to 0.01, using a multinomial scheduling strategy during training to reduce the learning rate to 0.001, setting the input image size to 224×224, and the batch size to 8.

3. The colon polyp image segmentation method based on cross-domain entropy-enhanced prototype contrastive learning according to claim 1, characterized in that, The pre-trained contrastive loss in step 3 , represented as: In the formula: Represents the projection features of the target domain Pixel features that are of the same category as the anchor point are considered positive samples. Projection features of the target domain Pixel features of different categories from the anchor point are considered negative samples. Indicates anchor point Compared with positive samples The similarity between them; Indicates anchor point With negative samples The similarity between them; It is the temperature coefficient.

4. The colon polyp image segmentation method based on cross-domain entropy-enhanced prototype contrastive learning according to claim 1, characterized in that, The data processing procedure of the entropy filtering module in step 4 is as follows: Step 4.1: Define each pixel Predicted entropy , represented as: In the formula: For pixels Belongs to the The predicted probability of a class; Total number of categories; Step 4.2: Pass the preset entropy threshold. Filtering a reliable set of pixels with high confidence , represented as: In the formula: Entropy threshold; A reliable set of pixels; Step 4.3: Utilize a reliable pixel set The Middle Calculation of the mean pixel feature of the class High-quality class prototypes , represented as: In the formula: For the first A reliable set of pixels; For pixels eigenvectors.

5. The colon polyp image segmentation method based on cross-domain entropy-enhanced prototype contrastive learning according to claim 1, characterized in that, The process of step 5 is as follows: Step 5.1, using pixel-level features As an anchor point, construct the category prototype. Guided pixel-level contrastive learning loss is applied and entropy-weighted optimization is performed. The specific process is as follows: Step 5.1.1: Using pixel-level features As anchor points, prototypes of the same category as the target are used as positive samples, and the rest of the prototype sets are used as positive samples. As negative samples, pixel-level comparison terms are constructed. , represented as: In the formula: For pixels eigenvectors; For pixels The prediction category; For the first The class's category prototype; For pixels The predicted category corresponds to the category prototype; It is the temperature coefficient; express and The similarity between them; for and The similarity between them; Step 5.1.2: Introduce an entropy-weighted strategy and define pixel weights. , represented as: In the formula: For pixels Entropy weights; For pixels Predicted entropy; Represents pixels Predicted entropy; Represents any pixel in the current batch; Represents the number of pixels in the current batch. The maximum value of the predicted entropy; Step 5.1.3: Calculate the entropy-weighted class-prototype contrast loss. , represented as: In the formula: This represents the set of pixels that participate in the contrastive learning process. Pixel-level contrast; Pixel weights; Step 5.2: Calculate the source domain supervised segmentation loss. , represented as: In the formula: Provide the actual segmentation label for the source domain; Represents cross-entropy loss; Indicates Dice loss; and They are respectively and Weighting coefficients; This is the source domain prediction probability map.

6. The colon polyp image segmentation method based on cross-domain entropy-enhanced prototype contrastive learning according to claim 1, characterized in that, Step 6 The value of is 0.1.

Images