Prototype-guided semi-supervised medical image segmentation method

Abstract

The application discloses a prototype-guided semi-supervised medical image segmentation method, which firstly introduces a class prototype to explicitly model the class conditional distribution center of each class, provides a class-level semantic anchor point for semi-supervised segmentation, expands pixel-level supervision into dual supervision combining pixel-level supervision and class-level structural constraint, and still can rely on the class-level semantic anchor point for discrimination when the pixel-level probability is unreliable, thereby solving the segmentation difficulty problem in the scene of blurred boundary and complex structure in the medical image, then based on the prototype similarity, the semantic of the pseudo label of the low confidence area is revised, the negative influence of the false label is reduced, finally, through the double-gated mechanism composed of the prediction confidence and the prototype similarity, reliable pixels are screened, the confirmation bias caused by the accumulation of false labels is inhibited, and the robustness of the model to geometric disturbance and light disturbance is further enhanced through multi-view consistency learning, and finally the segmentation accuracy and stability of the medical image are improved under a small amount of labeling.

Description

Technical Field

[0001] This invention relates to the field of medical imaging technology, and in particular to a prototype-guided semi-supervised medical image segmentation method. Background Technology

[0002] Medical image segmentation is a crucial foundational step in clinical auxiliary diagnosis, lesion quantification analysis, structural measurement, and efficacy evaluation. In recent years, deep learning methods have significantly improved the performance of medical image segmentation; however, mainstream methods typically rely on extensive pixel-level annotations. High-quality pixel-level annotation requires professional physicians to meticulously depict tissue or lesion boundaries pixel by pixel, resulting in high annotation costs and long processing times. Consequently, in real-world clinical scenarios, only a small amount of annotated data is often obtained, while a large amount of unannotated data remains underutilized.

[0003] Semi-supervised medical image segmentation aims to train a model using a small amount of labeled data and a large amount of unlabeled data, representing an important direction for solving the aforementioned problems. Most existing semi-supervised medical image training methods rely on pseudo-label learning and consistency regularization. Their main process involves the model first predicting the unlabeled image, and then using the high-confidence predictions as pseudo-labels for subsequent training. While these methods are effective in natural image tasks, they still have significant shortcomings in medical image scenarios.

[0004] Analysis revealed that the root cause of these shortcomings lies in the fact that existing methods are generally based on pixel-level probabilistic supervision, lacking explicit modeling of "category-level structural information." This results in models lacking reliable semantic anchors to evaluate and filter unsupervised signals. This lack of structural priors exposes a series of interconnected problems when facing the inherent characteristics of medical imaging, such as blurred boundaries, variable tissue morphology, and significant differences in imaging conditions: First, due to the lack of category-level structural constraints, the model makes decisions based solely on the probability values ​​of isolated pixels. Anatomical structures in medical images often possess relatively stable overall semantics and hierarchical relationships, such as retinal layers, optic cup / disc boundaries, and anterior segment boundaries. Relying solely on the probability values ​​of individual pixels is insufficient to stably describe these structures. When local pixel intensity fluctuates due to noise or individual differences, probability predictions are prone to oscillations, making it difficult to guarantee the continuity and anatomical plausibility of the segmentation results.

[0005] Second, when labels are scarce, false labels are prone to errors in areas with blurred boundaries, thin structures, low-contrast regions, or adjacent categories. Once erroneous false labels are incorporated into consistency learning, the model will continuously reinforce the error in subsequent iterations, thus forming a "confirmation bias," leading to training instability or even performance degradation.

[0006] Third, conventional methods for enhancing consistency typically only consider maintaining consistency between predictions across different views. However, if the predictions after geometric enhancement are not strictly aligned with the original space, additional misalignment errors may be introduced. Furthermore, if consistency constraints are also applied to unreliable pixels, semantic drift will be amplified.

[0007] Therefore, there is an urgent need for a semi-supervised medical image segmentation scheme that can explicitly utilize category-level structural priors, reduce the risk of false label noise propagation, and enhance the model's robustness to various perturbations. Summary of the Invention

[0008] To address the aforementioned problems, this invention proposes a prototype-guided semi-supervised medical image segmentation method. This method aims to solve the problem that existing semi-supervised medical image segmentation methods lack category-level structural priors as semantic anchors, resulting in the model's inability to effectively distinguish between reliable and unreliable unsupervised signals. This leads to a series of coupling problems such as pseudo-label noise propagation, confirmation bias, and consistency learning failure.

[0009] The technical solution of the present invention is as follows: A prototype-guided semi-supervised medical image segmentation method includes the following steps: Obtain a training dataset, which includes a labeled image set and an unlabeled image set, wherein the labeled images in the labeled image set correspond to pixel-level ground truth labels; Enhancement processing is performed on the unlabeled images in the unlabeled image set to generate an anchor point view and at least one enhanced view; The labeled image, the anchor point view, and the enhanced view are respectively input into the shared segmentation network to generate a pixel-level class probability map and an intermediate decoding feature map corresponding to each input image; For each segmentation category in the medical image segmentation task, a corresponding category prototype is constructed and maintained. The category prototype is used to represent the class conditional distribution center of the corresponding segmentation category in the feature space. Furthermore, the category prototype is dynamically updated based on the pixel features corresponding to the pixel-level true labels of the labeled image and the pixel features corresponding to the pixels in the anchor view that meet the preset confidence requirements. Initial pseudo-labels are generated based on the pixel-level category probability map of the anchor point view. Regions to be corrected in the initial pseudo-labels are determined. The similarity between each pixel feature in the intermediate decoded feature map and each category prototype is calculated. The predicted category of the region to be corrected is semantically corrected based on the similarity to obtain refined pseudo-labels. Pixels that simultaneously satisfy the prediction confidence condition and the prototype similarity condition are selected from the unlabeled images to form a reliable pixel set; At the pixel positions corresponding to the reliable pixel set, apply consistency constraints to the pixel-level category probability map of the enhanced view and the refined pseudo-labels corresponding to the anchor view, and calculate the unsupervised consistency loss. The supervised loss is calculated based on the pixel-level category probability map of the labeled image and the pixel-level ground truth label. The parameters of the shared segmentation network are updated based on the weighted sum of the supervised loss and the unsupervised consistency loss. The iteration stops when the preset convergence condition is met, and the trained segmentation model is output. The segmentation model is used to segment the target medical image to be segmented and generate the corresponding segmentation result map.

[0010] In some embodiments, the enhanced view includes a geometric view and / or a lighting view, wherein the geometric view is obtained by applying a geometric transformation to the unlabeled image, and the lighting view is obtained by applying a lighting transformation to the unlabeled image, wherein the geometric transformation includes at least one of rotation, flipping, cropping, and scaling, and the lighting transformation includes at least one of brightness adjustment, contrast adjustment, and color distribution adjustment.

[0011] In some implementations, the shared segmentation network is a segmentation network based on a convolutional neural network or an attention mechanism.

[0012] In some implementations, the category prototype is updated using an exponential moving average strategy, and the updated category prototype is a weighted sum of the historical prototype and the current batch mean features.

[0013] In some implementations, the similarity is a cosine similarity, and before calculating the similarity, the pixel features in the intermediate decoded feature map and the category prototype are respectively L2 normalized.

[0014] In some implementations, the prediction confidence condition is: the maximum predicted probability of the pixels in the unlabeled image is higher than a first preset threshold; the prototype similarity condition is: the similarity between the pixel features of the pixels in the unlabeled image and the prototype of the predicted category is higher than a second preset threshold.

[0015] In some implementations, the region to be corrected is a region where the prediction confidence is lower than a preset threshold, or a region where the predicted category is inconsistent with the category determined based on prototype similarity.

[0016] In some implementations, constructing and maintaining corresponding category prototypes for each segmentation category in the medical image segmentation task includes: maintaining a global category prototype for each segmentation category, or maintaining multiple sub-prototypes for segmentation categories with internal morphological differences greater than a preset degree, and selecting the sub-prototype with the highest similarity or performing weighted fusion of the similarity of each sub-prototype during semantic correction.

[0017] The beneficial effects of this invention are as follows: First, by introducing category prototypes to explicitly model the distribution centers of class conditions for each category, category-level semantic anchors are provided for semi-supervised segmentation. This extends pixel-level supervision to a dual supervision combining pixel-level supervision and category-level structural constraints, enabling the model to still make judgments using category-level semantic anchors even when pixel-level probabilities are unreliable. This effectively solves the problem of segmentation difficulties in scenarios with blurred boundaries and complex structures in medical images. Then, based on prototype similarity, pseudo-labels in low-confidence regions are semantically corrected to reduce the negative impact of erroneous labels. Finally, a dual-gating mechanism consisting of prediction confidence and prototype similarity is used to screen reliable pixels, suppressing confirmation bias caused by the accumulation of false label errors. Furthermore, multi-view... Figure 1 Consistency learning enhances the robustness of the model to geometric and illumination perturbations, ultimately improving the segmentation accuracy and stability of medical images with minimal annotation. Attached Figure Description

[0018] Figure 1 This is a schematic flowchart of the prototype-guided semi-supervised medical image segmentation method disclosed in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the content of this invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to this invention are shown in the accompanying drawings, not all of them.

[0020] This embodiment proposes a prototype-guided semi-supervised medical image segmentation method, such as... Figure 1 As shown, it includes the following steps: Step 1, Data Acquisition: Acquire the training dataset, which includes a labeled image set and an unlabeled image set. The labeled images in the labeled image set have pixel-level true labels.

[0021] In one example, the labeled dataset can be represented as D_l={(x_l, y_l)}, and the unlabeled dataset as D_u={x_u}, where x_l is the labeled image, y_l is the pixel-level ground truth label, and x_u is the unlabeled image. The labeled image set contains far fewer labeled samples than the unlabeled image set contains, by at least 10 times.

[0022] Step 2, View Construction: Enhance the unlabeled images in the unlabeled image set to generate an anchor view and at least one enhanced view.

[0023] In a preferred embodiment, the enhanced view includes a geometric view and / or a lighting view. The geometric view is obtained by applying a geometric transformation to the unlabeled image, and the lighting view is obtained by applying a lighting transformation to the unlabeled image. The geometric transformation includes at least one of rotation, flipping, cropping, and scaling, and the lighting transformation includes at least one of brightness adjustment, contrast adjustment, and color distribution adjustment. In this embodiment, the anchor point view undergoes weak enhancement processing. The enhanced view includes both a geometric view and a lighting view. These three views together constitute a multi-view input for subsequent anchor point spatial consistency learning in step 7.

[0024] In one example, for each unlabeled sample x_u, three views can be constructed: a weak view x_u^w, a geometric view x_u^geo, and an illumination view x_u^pho. The weak view x_u^w serves as an anchor space, used to generate pseudo-labels, update class prototypes, and as an alignment reference for subsequent consistency learning. The geometric view x_u^geo enhances the model's robustness to spatial perturbations, and the illumination view x_u^pho enhances the model's robustness to changes in imaging conditions.

[0025] Step 3, Feature Extraction: Input the labeled image, anchor view, and augmented view into the shared segmentation network to generate a pixel-level class probability map and intermediate decoding feature map for each input image.

[0026] In one example, the labeled image x_l, the anchor view, and the augmented view are input into the shared segmentation network f_θ (where θ represents the training parameters of the shared segmentation network), and the output is a pixel-level class probability map p and an intermediate decoded feature map z.

[0027] Optionally, the shared segmentation network described in this embodiment is a segmentation network based on a convolutional neural network or an attention mechanism. Specifically, the shared segmentation network can adopt a combination architecture of DeepLabV3+ and ResNet-50 encoder, or it can adopt a segmentation architecture based on U-Net, Attention U-Net, DPT, or Transformer. Any medical image segmentation network capable of outputting pixel-level class probability maps and intermediate decoded feature maps can be embedded in the technical framework of this embodiment; this embodiment is not limited to a specific segmentation backbone network.

[0028] Step 4, Prototype Maintenance: Construct and maintain corresponding category prototypes for each segmentation category in the medical image segmentation task. The category prototypes are used to represent the class conditional distribution center of the corresponding segmentation category in the feature space. Furthermore, the category prototypes are dynamically updated based on the pixel features corresponding to the pixel-level true labels of the labeled images and the pixel features corresponding to the pixels in the anchor view that meet the preset confidence requirements.

[0029] In one example, for each segmentation category in a medical image segmentation task, a corresponding category prototype is constructed and maintained. The total number of segmentation categories can be set to K. For any segmentation category c (1≤c≤K), its corresponding category prototype is denoted as p_c, which is used to characterize the class conditional distribution center of the segmentation category in the feature space.

[0030] Anatomical structures in medical images often possess relatively stable overall semantics and hierarchical relationships, such as retinal layers, optic cup / disc boundaries, and anterior segment boundaries. Relying solely on the probability value of a single pixel is insufficient to stably describe these structures. Therefore, this embodiment introduces category prototypes to explicitly model the class-conditional distribution centers of each category in the feature space, extending "pixel-level probabilistic supervision" to "pixel-level probabilistic supervision + category-level structural constraints," providing semantic anchors for subsequent pseudo-label correction and reliable region selection. In one example, the category prototype can be initialized as follows: For pixels belonging to each segmentation category in the labeled image set, the corresponding pixel features are collected, and the average of the pixel features for each category is calculated. This average is used as the initial category prototype for that category. For segmentation categories not appearing in the labeled samples, a random initialization method is used to establish their initial category prototypes.

[0031] In a preferred embodiment, the aforementioned category prototypes are updated using an exponential moving average strategy. The updated category prototype is a weighted average of the historical prototype and the current batch mean feature, i.e., updated category prototype = historical prototype × momentum coefficient + current batch mean feature × compensation coefficient. This approach balances the stability and adaptability of category prototype maintenance.

[0032] More preferably, when constructing and maintaining corresponding category prototypes for each segmentation category in a medical image segmentation task, the following two implementation methods can be adopted: For each segmentation category, maintain a global category prototype; or, for segmentation categories with internal morphological differences greater than a preset level, maintain multiple sub-prototypes, and select the sub-prototype with the highest similarity or perform weighted fusion of the similarities of each sub-prototype during semantic correction. When using a multi-sub-prototype strategy, the sub-prototype with the highest similarity is selected in subsequent semantic correction, and the similarities of each sub-prototype are weighted and fused to more accurately characterize categories with large internal morphological differences. For example, for categories with significant differences in lesion morphology or large boundary changes, a multi-prototype strategy can be used to improve segmentation accuracy.

[0033] Step 5, Pseudo-label generation and refinement: Generate initial pseudo-labels based on the pixel-level category probability map of the anchor point view, determine the regions to be corrected in the initial pseudo-labels, calculate the similarity between each pixel feature in the intermediate decoded feature map and each category prototype, and perform semantic correction on the predicted category of the region to be corrected based on the similarity to obtain refined pseudo-labels.

[0034] In a preferred embodiment, the similarity is cosine similarity. Furthermore, before calculating the similarity, the pixel features and category prototypes in the intermediate decoded feature map are L2 normalized respectively, so that the cosine similarity can more accurately measure the semantic closeness between the pixel features and the category prototypes.

[0035] Specifically, for any pixel feature z_i and class prototype p_c in the intermediate decoded feature map, the similarity between them can be expressed as s_c(z_i) = sim(z_i, p_c), where sim represents the cosine similarity operation. After calculating the similarity between the pixel feature and all class prototypes, the class corresponding to the maximum similarity is taken as the predicted class of the pixel prototype, i.e., predicted class = argmax_c s_c(z_i).

[0036] In another preferred embodiment, the region to be corrected is either a region where the prediction confidence is lower than a preset threshold, or a region where the predicted category is inconsistent with the category determined based on prototype similarity. Specifically, the former of the two preferred embodiments targets pixels with unstable network prediction probabilities, while the latter targets pixels with high prediction probabilities but inconsistent with the semantic structure of the category, using prototype similarity for correction.

[0037] Step 6, Reliable region selection: Pixels that simultaneously meet the prediction confidence condition and the prototype similarity condition are selected from the unlabeled images to form a reliable pixel set.

[0038] In a preferred embodiment, the prediction confidence condition is: the maximum predicted probability of pixels in the unlabeled image is higher than a first preset threshold; the prototype similarity condition is: the similarity between the pixel features of the unlabeled image and the prototype of the predicted category is higher than a second preset threshold. These two conditions together constitute a dual-gated risk control mechanism. The first gating is used to exclude pixels with unstable network probability distributions, and the second gating is used to exclude pixels with high probabilities but inconsistent with the semantic structure of the category. Only pixels that pass both gatings are included in the reliable pixel set for subsequent unsupervised loss calculation. The specific values ​​of the first and second preset thresholds can be adjusted according to the actual training data distribution, and those skilled in the art can determine them according to actual needs.

[0039] The aforementioned dual-gating design can simultaneously identify two types of risks: firstly, "pseudo-high confidence" risk pixels with high predicted probabilities but inconsistent with the semantics of the category structure; and secondly, "rescueable" pixels with moderate predicted probabilities but which can be corrected through prototypes. This mechanism not only suppresses false label confirmation bias but also improves the utilization efficiency of unlabeled data.

[0040] It should be noted that this embodiment involves multiple preset judgment conditions, each serving a different stage. Specifically: the preset confidence requirement in step 4 is used to filter pixel features participating in prototype updates; the preset threshold in step 5 is used to determine which regions need to be corrected; and the first and second preset thresholds in step 6 are used to construct a dual-gating mechanism to filter a reliable pixel set. The threshold values ​​of the above preset conditions are independent of each other and can be adjusted according to different training stages and different data distributions.

[0041] Step 7, Consistency Constraint: At the pixel positions corresponding to the reliable pixel set, apply consistency constraints to the pixel-level category probability map of the enhanced view and the refined pseudo-labels corresponding to the anchor view, and calculate the unsupervised consistency loss.

[0042] Specifically, for the geometric view, since geometric transformations change the spatial position of pixels, the prediction results of the geometric view need to be mapped back to the coordinate system of the anchor view through the corresponding inverse transformation before participating in the consistency loss calculation, to ensure that the supervision signal and the prediction results are aligned in the same anchor space. For the lighting view, since lighting transformations do not change the spatial position, its prediction results can be directly used for consistency constraints with the refined pseudo-labels at the corresponding positions of the anchor view.

[0043] Step 8, Model Optimization and Output: Calculate the supervised loss based on the pixel-level class probability map and pixel-level ground truth labels of the labeled image. Update the parameters of the shared segmentation network based on the weighted sum of the supervised loss and the unsupervised consistency loss. Stop the iteration when the preset convergence condition is met, and output the trained segmentation model. The segmentation model is used to segment the target medical image to be segmented and generate the corresponding segmentation result image.

[0044] In one example, the overall objective loss function can be expressed as L = L_x + λL_u, where L_x is the supervised segmentation loss, L_u is the unsupervised consistency loss, and λ is the weighting coefficient of the unsupervised loss, used to balance the contributions of supervised and unsupervised consistency losses in model optimization. Those skilled in the art can adjust the value of λ according to the actual training situation. For example, a smaller λ value can be set in the early stages of training to prioritize supervised learning, and the λ value can be gradually increased as training progresses to enhance the contribution of unsupervised signals.

[0045] Finally, it should be noted that the methods described above can be applied to various medical image segmentation scenarios. For example, in the OCT retinal layer segmentation scenario, the input image is a circumflex disc OCT image, and the segmentation categories include RNFL, GCIPL, Choroid, and background. After training, the model can output smooth and continuous retinal layer segmentation results for layer thickness measurement. In the fundus optic cup / optic disc segmentation scenario, the optic disc region ROI is first extracted using a coarse localization method, and then the ROI image is input into the method of this embodiment for semi-supervised training. The segmentation categories include optic cup, optic disc, and background. After training, clinical indicators such as the cup-to-disc ratio can be further calculated. In the AS-OCT anterior segment structure segmentation scenario, the segmentation targets include structures such as the lens, lens nucleus, left iris, right iris, and anterior chamber. After training, it can be used to assist in the automatic measurement of anterior segment structures and lesion assessment. Those skilled in the art can implement the methods of this embodiment based on the disclosure of this specification, combined with common deep learning training frameworks and conventional GPU server environments.

[0046] Those skilled in the art should know that the above-mentioned types of prototype maintenance and dynamic updates, multi-view Figure 1 Consistent learning and dual-gated risk control are performed only during the model training phase. During the inference phase, only the trained shared segmentation network needs to be retained. The target medical image to be segmented is input into the network to obtain the corresponding segmentation result image. There is no need to perform prototype maintenance, pseudo-label generation and refinement operations. Therefore, the deployment complexity and computational cost during the inference phase are comparable to those of conventional segmentation networks.

[0047] The above embodiments are merely illustrative of the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made based on the essence of the content of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A prototype-guided semi-supervised medical image segmentation method, characterized in that, Includes the following steps: Obtain a training dataset, which includes a labeled image set and an unlabeled image set, wherein the labeled images in the labeled image set correspond to pixel-level ground truth labels; Enhancement processing is performed on the unlabeled images in the unlabeled image set to generate an anchor point view and at least one enhanced view; The labeled image, the anchor point view, and the enhanced view are respectively input into the shared segmentation network to generate a pixel-level class probability map and an intermediate decoding feature map corresponding to each input image; For each segmentation category in the medical image segmentation task, a corresponding category prototype is constructed and maintained. The category prototype is used to represent the class conditional distribution center of the corresponding segmentation category in the feature space. Furthermore, the category prototype is dynamically updated based on the pixel features corresponding to the pixel-level true labels of the labeled image and the pixel features corresponding to the pixels in the anchor view that meet the preset confidence requirements. Initial pseudo-labels are generated based on the pixel-level category probability map of the anchor point view. Regions to be corrected in the initial pseudo-labels are determined. The similarity between each pixel feature in the intermediate decoded feature map and each category prototype is calculated. The predicted category of the region to be corrected is semantically corrected based on the similarity to obtain refined pseudo-labels. Pixels that simultaneously satisfy the prediction confidence condition and the prototype similarity condition are selected from the unlabeled images to form a reliable pixel set; At the pixel positions corresponding to the reliable pixel set, apply consistency constraints to the pixel-level category probability map of the enhanced view and the refined pseudo-labels corresponding to the anchor view, and calculate the unsupervised consistency loss. The supervised loss is calculated based on the pixel-level category probability map of the labeled image and the pixel-level ground truth label. The parameters of the shared segmentation network are updated based on the weighted sum of the supervised loss and the unsupervised consistency loss. The iteration stops when the preset convergence condition is met, and the trained segmentation model is output. The segmentation model is used to segment the target medical image to be segmented and generate the corresponding segmentation result map.

2. The prototype-guided semi-supervised medical image segmentation method according to claim 1, characterized in that, The enhanced view includes a geometric view and / or a lighting view. The geometric view is obtained by applying a geometric transformation to the unlabeled image, and the lighting view is obtained by applying a lighting transformation to the unlabeled image. The geometric transformation includes at least one of rotation, flipping, cropping, and scaling, and the lighting transformation includes at least one of brightness adjustment, contrast adjustment, and color distribution adjustment.

3. The prototype-guided semi-supervised medical image segmentation method according to claim 1, characterized in that, The shared segmentation network is a segmentation network based on convolutional neural networks or attention mechanisms.

4. The prototype-guided semi-supervised medical image segmentation method according to claim 1, characterized in that, The category prototype is updated using an exponential moving average strategy. The updated category prototype is composed of a weighted average of the historical prototype and the mean feature of the current batch.

5. The prototype-guided semi-supervised medical image segmentation method according to claim 1, characterized in that, The similarity is cosine similarity, and before calculating the similarity, the pixel features in the intermediate decoded feature map and the category prototype are respectively L2 normalized.

6. The prototype-guided semi-supervised medical image segmentation method according to claim 1, characterized in that, The prediction confidence condition is: the maximum predicted probability of the pixels in the unlabeled image is higher than a first preset threshold; the prototype similarity condition is: the similarity between the pixel features of the pixels in the unlabeled image and the prototype of the predicted category is higher than a second preset threshold.

7. The prototype-guided semi-supervised medical image segmentation method according to claim 1, characterized in that, The region to be corrected is a region where the prediction confidence is lower than a preset threshold, or a region where the predicted category is inconsistent with the category determined based on prototype similarity.

8. The prototype-guided semi-supervised medical image segmentation method according to claim 1, characterized in that, The construction and maintenance of corresponding category prototypes for each segmentation category in the medical image segmentation task includes: maintaining a global category prototype for each segmentation category, or maintaining multiple sub-prototypes for segmentation categories with internal morphological differences greater than a preset degree, and selecting the sub-prototype with the highest similarity or performing weighted fusion of the similarity of each sub-prototype during semantic correction.

Images