A medical image segmentation method and device based on domain adaptation and process control

Abstract

The application relates to a medical image segmentation method and device based on domain adaptation and process control, which comprises the following steps: S1, acquiring labeled and unlabeled medical image data, and building a semi-supervised segmentation network with a visual basic model as an encoding backbone and a teacher-student framework; S2, performing multi-scale refinement and field correction on features through a domain adaptive refinement (DAR) module to generate medical domain adaptive encoding features; S3, constructing an uncertainty weight map through an uncertainty-guided hierarchical consistency (UHC) module to constrain the consistency of multi-layer encoding and decoding features of the teacher-student; and S4, extracting class prototypes from the teacher branch through a prototype-guided feature refinement (PFR) module, and enhancing the decoding features of the student branch. The application can inhibit pseudo-label noise under the condition of few labels, improve the segmentation precision of complex medical targets, and is suitable for the field of medical image intelligent analysis.

Description

Technical Field

[0001] This invention relates to the field of medical image segmentation, and specifically to a medical image segmentation method and apparatus based on domain adaptation and process control. Background Technology

[0002] Medical image segmentation is a crucial step in computer-aided diagnosis, lesion localization, treatment planning, and efficacy evaluation. With the widespread application of medical imaging technologies such as CT, MRI, and ultrasound, a vast amount of medical image data has accumulated in clinical settings. However, high-quality pixel-level or voxel-level annotation typically requires professional physicians to draw the image frame by frame, resulting in high annotation costs and long processing times, thus limiting the scale of labeled data actually available for training.

[0003] To reduce reliance on manual annotation, semi-supervised segmentation methods are increasingly being adopted in existing technologies. These methods improve model performance by jointly utilizing a small amount of labeled data and a large amount of unlabeled data. Most existing semi-supervised medical image segmentation methods employ a teacher-student consensus learning framework. A teacher model generates pseudo-labels for unlabeled samples, and these pseudo-labels are then used to constrain the training of the student model. While this approach can alleviate the labeling problem to some extent, its performance is highly dependent on the quality of the pseudo-labels. When target boundaries are blurred, foreground-background contrast is low, or the structural morphology is complex, the pseudo-labels generated by the teacher model are prone to containing noise, which then propagates layer by layer during training, affecting segmentation accuracy and model stability.

[0004] On the other hand, visual foundation models possess strong general representation capabilities in natural image tasks, leading existing techniques to attempt to transfer them to medical image segmentation scenarios. However, natural images and medical images differ significantly in texture distribution, imaging mechanisms, structural morphology, and noise characteristics. Directly using visual foundation models to encode features can easily result in domain shift issues, making it difficult for the model to fully adapt to medical image features. Furthermore, existing techniques typically lack explicit modeling of high-confidence category semantics during the decoding stage, making it difficult to consistently provide stable and clean category priors for student branches under conditions of limited annotation.

[0005] In summary, existing technologies still have shortcomings in adapting visual basic models to the medical domain, suppressing noise propagation during training, and enhancing category semantics during the decoding stage. There is an urgent need for a new semi-supervised medical image segmentation method to improve the accuracy, stability, and generalization ability of medical image segmentation. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a medical image segmentation method and apparatus based on domain adaptation and process control. This method can effectively alleviate the domain shift problem when transferring the visual base model to medical images under conditions of low annotation, suppress the propagation of low-confidence pseudo-supervision during the training process, and enhance the category semantic expression in the decoding stage, thereby improving the accuracy and stability of medical image segmentation results.

[0007] Specifically, the method includes the following steps: S1: Obtain labeled and unlabeled medical image data, and construct a semi-supervised medical image segmentation network based on a teacher-student framework with a visual basic model as the encoding backbone. S2: The multi-level features output by the encoder are refined at multiple scales and corrected by the domain adaptive refinement module (DAR) to obtain medical domain-adaptive coding features; S3: Construct an uncertainty weight graph through an uncertainty-guided hierarchical consistency module (UHC) and impose consistency constraints on the multi-layer encoder features and multi-layer decoder features of the teacher branch and the student branch. S4: Extract the category prototype from the high-confidence region of the teacher branch through the prototype-guided feature refinement module (PFR), and inject the category prototype into the decoding features of the student branch to obtain the enhanced decoding features.

[0008] Preferably, S1 includes the following steps: S1.1: Obtain labeled and unlabeled medical image datasets; S1.2: Perform weak enhancement and strong enhancement processing on unlabeled medical images respectively, and perform conventional enhancement processing on labeled medical images; S1.3: Input the weakly enhanced unlabeled image into the teacher branch, and input the strongly enhanced unlabeled image into the student branch; S1.4: Construct a supervised segmentation loss using the annotation information corresponding to the labeled medical images, and construct a semi-supervised training process by combining it with unlabeled medical images.

[0009] Preferably, S2 includes the following steps: S2.1: Construct a domain-adaptive refining module; S2.2: Perform multi-scale local modeling on the encoder output features based on the domain adaptive refinement module to obtain local enhanced features; S2.3: Based on the learnable domain query vector in the domain adaptive refinement module, perform similarity matching and offset correction on the local enhanced features to obtain medical domain adapted coding features; Preferably, S2.2 includes the following steps: S2.2.1: By converting the output features of the i-th layer of the encoder into a spatial feature map, local response features under different receptive fields are extracted using 3×3, 5×5, and 7×7 depth convolutions, respectively. These local response features are then concatenated and fused using convolution to obtain local enhancement features.

[0010]

[0011]

[0012]

[0013]

[0014] In the above formula, This represents the output feature of the i-th layer encoder. This represents a 3×3 depthwise separable convolution. This represents a 5×5 depthwise separable convolution. This represents a 7×7 depthwise separable convolution. Indicates fusion characteristics, Represents a 1×1 convolution. Indicates matrix concatenation. Indicates local enhancement features; Preferably, S2.3 includes the following steps: S2.3.1: By setting a learnable domain query vector, a similarity matrix is ​​calculated with the local enhancement features, and a domain offset correction term is generated based on the similarity matrix to obtain the domain-adaptive refined encoded features:

[0015]

[0016]

[0017] In the above formula, Represents the similarity matrix. Indicates the first The learnable domain query vector corresponding to the layer. Indicates the feature channel dimension. Indicates the feature correction amount. This represents a multilayer perceptron. The output features are represented by the domain adaptive refinement.

[0018] Preferably, S3 includes the following steps: S3.1: Construct a pixel-level or voxel-level uncertainty weight map based on the teacher branch prediction results and the student branch prediction results; S3.2: Based on the mapped uncertainty weight graph, apply weighted consistency constraints to the multi-layer encoder features and multi-layer decoder features of the teacher branch and student branch; Preferably, S3.1 includes the following steps: S3.1.1: By performing probability normalization operations on the teacher branch prediction results and the student branch prediction results respectively, the teacher branch category probability map and the student branch category probability map are obtained. The average uncertainty weight map is then obtained by averaging the two.

[0019]

[0020]

[0021]

[0022] in, This represents the student branch prediction result. This indicates the teacher branch prediction results. Represents the numerically stable term. Represents the branch entropy of the student. Represents the branch entropy of the teacher. Represents the uncertainty sensitivity coefficient. Represents an uncertainty weighting graph; Preferably, S3.2 includes the following steps: S3.2.1: By resampling the uncertainty weight map to different level feature resolutions, weighted consistency constraints are applied to the features of the multi-layer encoder and the multi-layer decoder.

[0023]

[0024]

[0025] in, Represents weight mapping, Indicates the first Layer features, Indicates the first Approval Indicates the first One channel, Indicates the number of encoder layers. Indicates the number of decoder layers. Indicates encoder loss, This represents the decoder loss.

[0026] Preferably, S4 includes the following steps: S4.1: Calculate the category confidence score based on the teacher branch prediction results and filter out high-confidence regions; S4.2: Extract category prototypes based on the teacher branch decoding features corresponding to the high-confidence region; S4.3: Perform attention interaction between the category prototype and the student branch decoding features to obtain the enhanced decoding features.

[0027] Preferably, S4.1 includes the following steps: S4.1.1: Calculate the category confidence score based on the teacher branch prediction results, and use the confidence score threshold to filter high-confidence regions:

[0028] in, Represents a confidence map for each pixel; Preferably, S4.2 includes the following steps: S4.2.1: By selecting the top-k response positions within the high-confidence region and performing weighted aggregation on the corresponding teacher branch decoding features, the category prototype is obtained:

[0029] in, For indicator functions Indicates the confidence threshold. Indicates a mask. The top-k proportionality coefficient. Indicates the preceding One response location, This indicates the weight corresponding to the corresponding position. Represents the category prototype; Preferably, S4.3 includes the following steps: S4.3.1: By normalizing and linearly mapping the category prototypes, and then interacting with the mapped category prototypes through attention, the enhanced student branch decoding features are obtained:

[0030]

[0031]

[0032] in, The decoding features represent the student branches. For the projection dimension, Represents the prototype features. This indicates a prototype-enhanced decoding feature.

[0033] The technical solution adopted in this invention is: a semi-supervised medical image segmentation device based on domain adaptation and process control, comprising: Data processing module: used to acquire labeled and unlabeled medical image data, and to construct input data for teacher and student branches; Domain Adaptive Refinement Module: Used to perform multi-scale local refinement and domain query correction on the multi-layer features output by the encoder to obtain medical domain-adaptive encoded features; Uncertainty-guided hierarchical consistency module: used to construct an uncertainty weight graph and apply weighted consistency constraints to the multi-layer encoder features and multi-layer decoder features of the teacher and student branches; Prototype-guided feature refining module: used to extract category prototypes from high-confidence regions of the teacher branch and semantically enhance the decoded features of the student branch.

[0034] The beneficial effects of the method of this invention are as follows: This invention acquires labeled and unlabeled medical image data and constructs a semi-supervised medical image segmentation network based on a teacher-student framework, with a visual basic model as the encoding backbone. Next, a domain-adaptive refinement module is used to perform multi-scale local refinement and domain query correction on the multi-layer features output by the encoder, obtaining medical domain-adaptive encoding features. Then, an uncertainty weight map is constructed through an uncertainty-guided hierarchical consistency module, and consistency constraints are applied to the multi-layer encoder features and multi-layer decoder features of the teacher and student branches, thereby suppressing the propagation of low-confidence pseudo-supervision during training. Furthermore, a prototype-guided feature refinement module extracts category prototypes from the high-confidence regions of the teacher branch and injects these category prototypes into the decoded features of the student branch to enhance category semantic expression. Finally, the enhanced decoded features are input into the segmentation head to generate the medical image segmentation result. Compared with existing technologies, this invention can, on the one hand, reduce the domain offset effect when the visual base model is directly transferred to medical image scenes and improve the ability of encoded features to express medical structures; on the other hand, it can improve the stability of the semi-supervised training process through uncertainty constraints and high-confidence prototype enhancement mechanisms, reduce error accumulation caused by pseudo-label noise, and ultimately improve the segmentation accuracy and reliability of blurred boundary regions, low-contrast regions and complex target regions. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention, the following is a brief introduction to the drawings used in the prior art and embodiments. The following drawings are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1This is a flowchart illustrating a medical image segmentation method based on domain adaptation and process control according to the present invention. Figure 2 This is a schematic diagram of the model architecture of a medical image segmentation method based on domain adaptation and process control according to the present invention; Figure 3 This is a schematic diagram of the domain adaptive refining module structure of a medical image segmentation method based on domain adaptation and process control according to the present invention. Figure 4 This is a schematic diagram of the uncertainty-guided hierarchical consistency module structure of a medical image segmentation method based on domain adaptation and process control according to the present invention. Figure 5 This is a schematic diagram of the prototype-guided feature refinement module structure of a medical image segmentation method based on domain adaptation and process control according to the present invention. Detailed Implementation Plan To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0037] This invention provides a medical image segmentation method and apparatus based on domain adaptation and process control, which can efficiently and accurately segment target regions in medical images with a small number of labeled samples, thereby improving the accuracy and efficiency of intelligent medical image analysis.

[0038] A typical embodiment of the present invention, taking the ACDC cardiac magnetic resonance imaging dataset, the BUSI breast ultrasound imaging dataset, or the UDIAT breast ultrasound imaging dataset as an example, refers to... Figure 1 The method includes the following steps: S1: Obtain labeled and unlabeled medical image data, and construct a semi-supervised medical image segmentation network based on a teacher-student framework with a visual basic model as the encoding backbone. S2: The multi-level features output by the encoder are refined at multiple scales and corrected by the domain adaptive refinement module (DAR) to obtain medical domain-adaptive coding features; S3: Construct an uncertainty weight graph through an uncertainty-guided hierarchical consistency module (UHC), and impose consistency constraints on the multi-layer encoder features and multi-layer decoder features of the teacher branch and student branch. S4: Extract the category prototype from the high-confidence region of the teacher branch through the prototype-guided feature refinement module (PFR), and inject the category prototype into the decoding features of the student branch to obtain the enhanced decoding features.

[0039] The following is a detailed description of each step.

[0040] Furthermore, as a preferred embodiment of this method, S1 includes the following steps: S1.1: Obtain labeled and unlabeled medical image datasets; Furthermore, the step of obtaining the labeled medical image dataset and the unlabeled medical image dataset specifically includes: Obtain medical image samples corresponding to the training task, and divide the samples with pixel-level or voxel-level annotations into labeled medical image datasets, and divide the unlabeled samples into unlabeled medical image datasets. S1.2: Perform weak enhancement and strong enhancement processing on unlabeled medical images respectively, and perform conventional enhancement processing on labeled medical images; Furthermore, the weak enhancement processing includes one or more of random flipping, random cropping, or intensity perturbation, and the strong enhancement processing includes one or more of intensity perturbation, contrast perturbation, noise perturbation, blur perturbation, occlusion perturbation, and geometric perturbation. S1.3: Input the weakly enhanced unlabeled image into the teacher branch, and input the strongly enhanced unlabeled image into the student branch; S1.4: Construct a supervised segmentation loss using the annotation information corresponding to the labeled medical images, and construct a semi-supervised training process by combining it with unlabeled medical images.

[0041] Furthermore, refer to Figure 3 S2 includes the following steps: S2.1: Construct the Domain Adaptive Refinement Module (DAR); S2.2: Perform multi-scale local modeling on the encoder output features based on the domain adaptive refinement module to obtain local enhanced features; S2.3: Based on the learnable domain query vector in the domain adaptive refinement module, perform similarity matching and offset correction on the local enhanced features to obtain medical domain adapted coding features.

[0042] Furthermore, S2.2 includes the following steps: S2.2.1: By converting the output features of the i-th layer of the encoder into a spatial feature map, local response features under different receptive fields are extracted using 3×3, 5×5, and 7×7 depth convolutions, respectively. These local response features are then concatenated and fused using convolution to obtain local enhancement features.

[0043]

[0044]

[0045]

[0046]

[0047] In the above formula, This represents the output feature of the i-th layer encoder. This represents a 3×3 depthwise separable convolution. This represents a 5×5 depthwise separable convolution. This represents a 7×7 depthwise separable convolution. Indicates fusion characteristics, Represents a 1×1 convolution. Indicates matrix concatenation. Indicates local enhancement features; Furthermore, S2.3 includes the following steps: S2.3.1: By setting a learnable domain query vector, a similarity matrix is ​​calculated with the local enhancement features, and a domain offset correction term is generated based on the similarity matrix to obtain the domain-adaptive refined encoded features:

[0048]

[0049]

[0050] In the above formula, Represents the similarity matrix. Indicates the first The learnable domain query vector corresponding to the layer. Indicates the feature channel dimension. Indicates the feature correction amount. This represents a multilayer perceptron. The output features are represented by the domain adaptive refinement.

[0051] Furthermore, refer to Figure 4 S3 includes the following steps: S3.1: Construct a pixel-level or voxel-level uncertainty weight map based on the teacher branch prediction results and the student branch prediction results; S3.2: Based on the mapped uncertainty weight graph, apply weighted consistency constraints to the multi-layer encoder features and multi-layer decoder features of the teacher branch and student branch; Furthermore, S3.1 includes the following steps: S3.1.1: By performing probability normalization operations on the teacher branch prediction results and the student branch prediction results respectively, the teacher branch category probability map and the student branch category probability map are obtained. The average uncertainty weight map is then obtained by averaging the two.

[0052]

[0053]

[0054]

[0055] in, This represents the student branch prediction result. This indicates the teacher branch prediction results. Represents the numerically stable term. Represents the branch entropy of the student. Represents the branch entropy of the teacher. Represents the uncertainty sensitivity coefficient. Represents an uncertainty weighting graph; Furthermore, S3.2 includes the following steps: S3.2.1: By resampling the uncertainty weight map to different level feature resolutions, weighted consistency constraints are applied to the features of the multi-layer encoder and the multi-layer decoder.

[0056]

[0057]

[0058] in, Represents weight mapping, Indicates the first Layer features, Indicates the first Approval Indicates the first One channel, Indicates the number of encoder layers. Indicates the number of decoder layers. Indicates encoder loss, This represents the decoder loss.

[0059] Furthermore, refer to Figure 5 S4 includes the following steps: S4.1: Calculate the category confidence score based on the teacher branch prediction results and filter out high-confidence regions; S4.2: Extract category prototypes based on the teacher branch decoding features corresponding to the high-confidence region; S4.3: Perform attention interaction between the category prototype and the student branch decoding features to obtain the enhanced decoding features.

[0060] Furthermore, S4.1 includes the following steps: S4.1.1: Calculate the category confidence score based on the teacher branch prediction results, and use the confidence score threshold to filter high-confidence regions:

[0061] in, Represents a confidence map for each pixel; Furthermore, S4.2 includes the following steps: S4.2.1: By selecting the top-k response positions within the high-confidence region and performing weighted aggregation on the corresponding teacher branch decoding features, the category prototype is obtained:

[0062] in, For indicator functions Indicates the confidence threshold. Indicates a mask. The top-k proportionality coefficient. Indicates the preceding One response location, This indicates the weight corresponding to the corresponding position. Represents the category prototype; Furthermore, S4.3 includes the following steps: S4.3.1: By normalizing and linearly mapping the category prototypes, and then interacting with the mapped category prototypes through attention, the enhanced student branch decoding features are obtained:

[0063]

[0064]

[0065] in, The decoding features represent the student branches. For the projection dimension, Represents the prototype features. This indicates a prototype-enhanced decoding feature.

[0066] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make other equivalent modifications or substitutions without departing from the spirit of the invention. These equivalent modifications or substitutions are included within the scope defined by the claims.

Claims

1. A medical image segmentation method based on domain adaptation and process control, characterized in that, The method includes the following steps: S1: Obtain labeled and unlabeled medical image data, and construct a semi-supervised medical image segmentation network based on a teacher-student framework with a visual basic model as the encoding backbone. S1 includes the following steps: S1.1: Obtain labeled and unlabeled medical image datasets; S1.2: Perform weak enhancement and strong enhancement processing on unlabeled medical images respectively, and perform conventional enhancement processing on labeled medical images; S1.3: Input the weakly enhanced unlabeled image into the teacher branch, and input the strongly enhanced unlabeled image into the student branch; S1.4: Construct a supervised segmentation loss using the annotation information corresponding to the labeled medical images, and construct a semi-supervised training process by combining it with unlabeled medical images; S2: Perform multi-scale local refinement and domain query correction on the multi-layer features output by the encoder through the domain adaptive refinement module to obtain medical domain adapted coding features; S2 includes the following steps: S2.1: Construct a domain-adaptive refining module; S2.2: Perform multi-scale local modeling on the encoder output features based on the domain adaptive refinement module to obtain local enhanced features; S2.3: Based on the learnable domain query vector in the domain adaptive refinement module, perform similarity matching and offset correction on the local enhanced features to obtain medical domain adapted coding features; S3: By using an uncertainty-guided hierarchical consistency module, consistency constraints are applied to the multi-level characteristics of the teacher branch and the student branch to obtain process control results; S3 includes the following steps: S3.1: Construct a pixel-level or voxel-level uncertainty weight map based on the teacher branch prediction results and the student branch prediction results; S3.2: Based on the mapped uncertainty weight graph, apply weighted consistency constraints to the multi-layer encoder features and multi-layer decoder features of the teacher branch and student branch; S4: Extract the category prototype from the high-confidence region of the teacher branch through the prototype-guided feature refinement module, and inject the category prototype into the decoding feature of the student branch to obtain the enhanced decoding feature; S4 includes the following steps: S4.1: Calculate the category confidence score based on the teacher branch prediction results and filter out high-confidence regions; S4.2: Extract category prototypes based on the teacher branch decoding features corresponding to the high-confidence region; S4.3: Perform attention interaction between the category prototype and the student branch decoding features to obtain the enhanced decoding features.

2. The medical image segmentation method and apparatus based on domain adaptation and process control according to claim 1, characterized in that, S2.2 includes the following steps: S2.2.1: By converting the output features of the i-th layer of the encoder into a spatial feature map, local response features under different receptive fields are extracted using 3×3, 5×5, and 7×7 depth convolutions, respectively. These local response features are then concatenated and fused using convolution to obtain local enhancement features. In the above formula, This represents the output feature of the i-th layer encoder. This represents a 3×3 depthwise separable convolution. This represents a 5×5 depthwise separable convolution. This represents a 7×7 depthwise separable convolution. Indicates fusion characteristics, Represents a 1×1 convolution. Indicates matrix concatenation. Indicates local enhancement features; S2.3 includes the following steps: S2.3.1: By setting a learnable domain query vector, a similarity matrix is ​​calculated with the local enhancement features, and a domain offset correction term is generated based on the similarity matrix to obtain the domain-adaptive refined encoded features: In the above formula, Represents the similarity matrix. Indicates the first The learnable domain query vector corresponding to the layer. Indicates the feature channel dimension. Indicates the feature correction amount. This represents a multilayer perceptron. The output features are represented by the domain adaptive refinement.

3. The medical image segmentation method and apparatus based on domain adaptation and process control according to claim 1, characterized in that, S3.1 includes the following steps: S3.1.1: By performing probability normalization operations on the teacher branch prediction results and the student branch prediction results respectively, the teacher branch category probability map and the student branch category probability map are obtained. The average uncertainty weight map is then obtained by averaging the two. in, This represents the student branch prediction result. This indicates the teacher branch prediction results. Represents the numerically stable term. Represents the branch entropy of the student. Represents the branch entropy of the teacher. Represents the uncertainty sensitivity coefficient. Represents an uncertainty weighting graph; S3.2 includes the following steps: S3.2.1: By resampling the uncertainty weight map to different level feature resolutions, weighted consistency constraints are applied to the features of the multi-layer encoder and the multi-layer decoder. in, Represents weight mapping, Indicates the first Layer features, Indicates the first Approval Indicates the first One channel, Indicates the number of encoder layers. Indicates the number of decoder layers. Indicates encoder loss, This represents the decoder loss.

4. The medical image segmentation method and apparatus based on domain adaptation and process control according to claim 1, characterized in that, S4.1 includes the following steps: S4.1.1: Calculate the category confidence score based on the teacher branch prediction results, and use the confidence score threshold to filter high-confidence regions: in, Represents a confidence map for each pixel; S4.2 includes the following steps: S4.2.1: By selecting the top-k response positions within the high-confidence region and performing weighted aggregation on the corresponding teacher branch decoding features, the category prototype is obtained: in, For indicator functions Indicates the confidence threshold. Indicates a mask. The top-k proportionality coefficient. Indicates the preceding One response location, This indicates the weight corresponding to the corresponding position. Represents the category prototype; S4.3 includes the following steps: S4.3.1: By normalizing and linearly mapping the category prototypes, and then interacting with the mapped category prototypes through attention, the enhanced student branch decoding features are obtained: in, The decoding features represent the student branches. For the projection dimension, Represents prototype features. This indicates a prototype-enhanced decoding feature.

5. A medical image segmentation device based on domain adaptation and process control, characterized in that, include: Data processing module: used to acquire labeled and unlabeled medical image data, and to construct input data for teacher and student branches; Domain Adaptive Refinement Module: Used to perform multi-scale local refinement and domain query correction on the multi-layer features output by the encoder to obtain medical domain-adaptive encoded features; Uncertainty-guided hierarchical consistency module: used to construct an uncertainty weight graph and apply weighted consistency constraints to the multi-layer encoder features and multi-layer decoder features of the teacher and student branches; Prototype-guided feature refining module: used to extract category prototypes from high-confidence regions of the teacher branch and semantically enhance the decoded features of the student branch.

Images