A semi-supervised medical image segmentation method based on uncertainty perception and feature prediction double-level collaborative alignment

Abstract

The application discloses a semi-supervised medical image segmentation method based on uncertainty perception and feature prediction double-level collaborative alignment, first constructs a medical image segmentation network containing a single feature encoder and a multi-scale auxiliary prediction head; secondly, weak data disturbance of basic space transformation and strong data disturbance based on mask shielding are applied to the unlabeled images respectively; then, based on the multi-scale probability graph output by the weak disturbance, the information entropy distribution fusion weight is calculated to generate the consensus pseudo label, and the variance is estimated to construct the consistency constraint of uncertainty correction; meanwhile, the network internal random inactivation module is closed to predict the strong disturbance image, a feature prediction double-level collaborative alignment strategy is introduced, an asymmetric prototype contrast distillation loss is constructed by cutting off the weak disturbance feature gradient at the deep feature level, a distribution alignment loss is constructed by using bidirectional KL divergence at the output level, and a unified double-level feature-distribution alignment loss is constructed through the double-level collaborative constraint of "feature-distribution".

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and medical image processing, and relates to a computer vision method based on deep learning. More specifically, this invention relates to a semi-supervised medical image segmentation method based on a two-level collaborative alignment of uncertainty perception and feature prediction, which is particularly suitable for high-precision and robust automatic segmentation of two-dimensional and three-dimensional medical images in scenarios where labeled data is extremely scarce. Background Technology

[0002] In recent years, deep learning technologies, represented by convolutional neural networks, have made groundbreaking progress in the field of medical image segmentation. However, the fully supervised learning paradigm is highly dependent on massive amounts of high-quality labeled data. In the field of 3D medical imaging, delineating organs or lesions layer by layer is extremely time-consuming and labor-intensive, and faces serious privacy restrictions. To overcome the data bottleneck, semi-supervised learning has emerged.

[0003] Although existing semi-supervised methods (such as those based on consistency regularization) have achieved some success, the following technical bottlenecks still exist when dealing with complex three-dimensional medical anatomy structures: First, there's the challenge of confirmation bias and uncertainty in local edges. In standard semi-supervised learning paradigms, models heavily rely on pseudo-labels generated from unlabeled data as supervisory signals to guide subsequent training. However, existing conventional methods typically employ a one-size-fits-all approach, treating all voxel regions in an image equally. In complex medical images, due to partial volume effects or extremely low contrast between tissues, lesion edges and transitional areas of anatomical structures are often extremely blurred. In these challenging regions, models are not only prone to errors but can even produce "high-confidence false predictions." If these heavily noisy predictions are directly incorporated into the next round of training as absolute "true labels," the erroneous signals will continuously backpropagate and accumulate within the network. This phenomenon leads the model into a quagmire of "confirmation bias," which not only amplifies prediction errors but also severely erodes the model's ability to generalize to unknown data, ultimately resulting in the degradation of segmentation boundaries and the adhesion of anatomical structures.

[0004] Second, traditional data perturbation strategies struggle to cope with the complex and ever-changing artifacts in clinical imaging. Existing semi-supervised medical image segmentation methods often rely on basic spatial geometric transformations (such as random flipping and affine transformations) or simple pixel-level noise additions when constructing data consistency constraints. However, these mild transformations are simply incapable of effectively simulating the severe tissue occlusion, metallic artifacts, motion artifacts, and local visual field defects that frequently occur in real clinical medical scans. This results in the model only learning shallow texture features under ideal conditions during training, failing to truly grasp the global topological structure and contextual dependencies of organs in three-dimensional space. When faced with highly heterogeneous and severely degraded real-world clinical inputs, models trained with weak perturbations often prove extremely fragile, lacking robustness and prone to large-scale misclassification or underclassification.

[0005] Third, the vulnerability and oscillations of the high-dimensional decision boundary of the model under implicit stochastic mechanisms. Medical images contain anatomical structures with extremely high spatial complexity and the curse of dimensionality, requiring segmentation networks to possess extremely accurate and stable discrimination capabilities. However, to prevent overfitting, deep learning models typically incorporate implicit stochastic mechanisms such as random deactivation or feature sampling. When faced with unlabeled data, this internal uncertainty can easily lead to drastic fluctuations in the probability output distribution generated by the network during multiple forward propagations for the same input image. This unconstrained internal output oscillation makes the classification decision boundary depicted by the network in the high-dimensional feature space extremely uneven and rough. Even the slightest perturbation or noise in the input image can severely shift the already fragile decision boundary, leading to local collapse of the prediction results and greatly limiting the stability and coherence of semi-supervised ensembles.

[0006] Fourth, the severe representational shift and semantic misalignment dilemma in the high-dimensional feature latent space. In advanced semi-supervised paradigms, to enhance the model's robustness, highly destructive data perturbations are often applied to unlabeled data. While this forces the model to learn harder samples, when severely damaged images are input into the network, the high-dimensional features extracted at deep bottlenecks often undergo severe manifold shifts, deviating significantly from the true anatomical semantic distribution formed by clean, complete data. In this case, relying solely on pixel-level probability maps at the network's final output layer for consistency constraints, while ignoring the alignment of feature vectors in the deep latent space, is merely a superficial solution. This spatial disconnect between deep features and shallow outputs not only makes it difficult to guarantee global semantic consistency but also easily leads to feature collapse, causing the model to fall into local overfitting on strongly perturbed branches and lose its cross-view generalization ability.

[0007] In response to the four major pain points and technical bottlenecks currently facing the field of semi-supervised medical image segmentation, there is an urgent need for a new architecture that can simultaneously take into account uncertainty perception, robustness against strong interference, and accurate alignment of deep features. This invention is proposed based on this urgent clinical and algorithmic need. Summary of the Invention

[0008] The purpose of this invention is to overcome the problems of pseudo-label confirmation bias, severe artifact interference, and deep feature space representation shift in existing semi-supervised medical image segmentation techniques, and to provide a semi-supervised medical image segmentation method based on a two-level collaborative alignment of uncertainty perception and feature prediction. This method achieves the construction of high-purity consensus pseudo-labels and dynamic uncertainty correction by introducing multi-scale prediction adaptive fusion based on information entropy; it achieves vertical multi-dimensional collaborative constraints between the deep representation space and the shallow decision boundary through an innovative two-level collaborative alignment strategy for feature prediction; and it constructs a hierarchical architecture spanning the "decoder deep latent space features" and the "decoder shallow prediction output distribution," using the pure deep prototype under weak perturbation and the stable decision distribution as two-level anchor points to collaboratively constrain the uncertain data flow under strong perturbation and random deactivation. This strategy not only effectively prevents the representation collapse and spatial shift of high-dimensional features at the bottom level, but also significantly enhances the continuity and smoothness of the model decision boundary at the top level, greatly strengthening the network's unsupervised robust perception capability against spatial occlusion and noise interference from the vertical topological dimension.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A semi-supervised medical image segmentation method based on two-level collaborative alignment of uncertainty perception and feature prediction includes the following steps: Step 1: Provide labeled medical image datasets and unlabeled medical image datasets, and construct a multi-scale pyramid segmentation network and internal state control mechanism: Initialize the medical image segmentation network, which includes an encoder, a decoder, a main prediction head, multiple auxiliary prediction heads, and a random deactivation module; Step 2, Constructing a Hybrid Strong and Weak Perturbation Input Stream: Sample images from the labeled and unlabeled datasets to construct the current batch. Apply weak and strong data perturbations of different external interference intensities to the unlabeled images in the current batch to obtain weak data perturbation images and strong data perturbation images, thus constructing an asymmetric distributed training input. Step 3: Input the weak data perturbation image into the medical image segmentation network, extract deep representation features, and generate a multi-scale weak perturbation prediction probability map from the output branch; Step 4: Calculate the information entropy of each scale based on the weak perturbation prediction probability map, allocate fusion weights through information entropy and generate consensus pseudo-labels by weighted averaging; combine the prediction results of each scale with the KL divergence of the consensus pseudo-labels to estimate the uncertainty variance and construct the uncertainty-corrected multi-scale consensus loss. Step 5: Turn off the random deactivation module, input the image with strong data perturbation into the medical image segmentation network, and obtain the strong perturbation prediction probability map and the deep representation features of the strong perturbation. Step 6: Align the deep representation features with strong data perturbation and the deep representation features with weak data perturbation at the latent space feature level, and align the decision boundary distribution of the strong perturbation prediction probability map and the weak perturbation master prediction probability map to construct a unified hierarchical architecture feature-distribution alignment loss. Through this step, a feature-prediction co-alignment (FPA) strategy that spans deep features and shallow outputs is constructed.

[0010] Step 7: Combine the depth supervision loss of labeled images (including cross-entropy loss and Dice loss), the uncertainty-corrected multi-scale consistency loss, and the hierarchical architecture feature-distribution alignment loss to construct the overall objective function, update the network parameters, and repeat steps 2 to 7 until the model converges. Output the trained network for medical image segmentation.

[0011] In one embodiment, step 1, in order to eliminate the heterogeneity of data collected by different devices, performs standardized preprocessing on all image data, including resampling the image to a uniform voxel spacing, performing random rotation and random cropping data enhancement, and may further truncate and normalize the pixel grayscale values.

[0012] In one embodiment, the medical image segmentation network is a single-network structure implemented based on a multi-scale pyramid segmentation network. It employs an encoder-decoder structure, including a single feature encoder for extracting multi-scale deep semantic features and a single feature decoder for resolution restoration. The main prediction head is located at the final output layer of the single feature decoder, and an auxiliary prediction head is attached to different feature scale levels of the single feature decoder, collectively constructing a multi-scale pyramid prediction architecture. The medical image segmentation network includes a random deactivation module and state control variables for controlling the operation of the random deactivation module. Structurally, the random deactivation module is a Dropout-like random deactivation layer located within the medical image segmentation network. Functionally, it is used to randomly mask some features in the input feature map according to a preset deactivation probability when enabled, and to set the deactivation probability to zero when disabled, thereby introducing internal random perturbation and reducing overfitting during network training. In the training process of this invention, the operating state of the random deactivation module is changed by adjusting the state control variable. This ensures the random deactivation module is enabled during forward propagation of images with weak data perturbations and disabled during forward propagation of images with strong data perturbations, achieving dual asymmetric forward propagation within the same training batch. Specifically, during the first forward propagation of images with weak data perturbations, the random deactivation module remains enabled and randomly masks some features according to a preset deactivation probability. Before the second forward propagation of images with strong data perturbations, the deactivation probability of the random deactivation module is set to zero, effectively disabling it and preventing the simultaneous superposition of strong data perturbations and random deactivation perturbations within the network. Both forward propagations use the same medical image segmentation network and share network parameters. The weak perturbation prediction results and deep representation features obtained from the first forward propagation, and the strong perturbation prediction results and deep representation features obtained from the second forward propagation, are used for subsequent output-level distribution alignment and feature-level comparative distillation, respectively.

[0013] In one embodiment, step 2 involves retaining standard data augmentation as a weak data perturbation image for the current batch of images, and randomly generating occlusion blocks of a preset size based on a mask occlusion strong perturbation strategy and applying them to the current batch of images to generate a corresponding strong data perturbation image. Specifically: Weak data perturbation: Preserving the spatial transformation of standard data augmentation and maintaining the anatomical semantic integrity of the image, the weakly perturbed image is input into a network under a normal random deactivation state for the first forward propagation. The standard data augmentation includes random rotation and random cropping. Strong data perturbation: Construct a mask tensor with all 1s of the exact same size as the input 3D medical image patch; randomly select a center coordinate within the mask tensor to generate a 2D pixel block or 3D voxel block of a preset size, and forcibly set the values ​​within this pixel block or voxel block to zero, obtaining the target mask tensor; multiply the current batch of unlabeled medical images element-wise with the target mask tensor, thereby erasing the values ​​of the corresponding regions in the images to zero, simulating real clinical lesion occlusion or artifact interference. Finally, input the strongly perturbed images into the network for a second forward propagation.

[0014] This invention achieves pyramid uncertainty perception and adaptive generation of consensus pseudo-labels through steps 3 to 5. The network is used for the first forward propagation of the weakly perturbated image, and the main prediction head and each auxiliary prediction head output multi-scale weakly perturbated prediction probability maps. (in ).

[0015] In one embodiment, step 4, generating consensus pseudo-labels, involves: first, obtaining multi-scale weakly perturbated prediction probability maps output by the main prediction head and each auxiliary prediction head; and then, using an adaptive weighting function based on information entropy, independently calculating the information entropy of each prediction probability map at the pixel or voxel level. in For the number of semantic categories, To prevent the use of tiny constants in division by zero, an adaptive weighting function based on a negative exponential mapping is employed. Dynamic transformation and fusion weights. The adaptive weight function is established through a negative exponential mapping, such that the smaller the information entropy at a given scale, the closer the fusion weight assigned to that scale is to 1; and the larger the information entropy, the closer the fusion weight assigned to that scale is to 0.

[0016] Finally, the multi-scale weak perturbation prediction probability maps are weighted, averaged, and normalized according to the fusion weights to generate high-purity consensus pseudo-labels. In one embodiment, step 4, constructing the multi-scale consensus loss method with uncertainty correction, is as follows: For the probability maps output by the main prediction head and each auxiliary prediction head, calculate the KL divergence between them and the consensus pseudo-label, and sum the divergences of all scales as the multi-scale prediction variance of the current batch. The uncertainty penalty term is constructed using the negative exponent of the multi-scale prediction variance. This penalty term is then used as a weight to dynamically adjust the mean square error between each scale's prediction and the consensus pseudo-label, thereby automatically suppressing noise interference at low-contrast and blurred boundary areas. Specifically, this step constructs the uncertainty-corrected multi-scale consensus loss using the following formula. : Where K is the total number of output heads, and an uncertainty penalty term is used. Automatically suppress noise at blurred anatomical boundaries.

[0017] In one embodiment, step 5, disabling the random deactivation module, specifically involves: before the second forward propagation, by setting a state control variable, forcing the deactivation probability of the random deactivation module integrated across all network layers in the medical image segmentation network to be zero, thereby eliminating the explicit randomness of the network's internal structure and ensuring that the strongly perturbed input stream can pass through the complete feature manifold space without loss. Subsequently, the strongly perturbed image is input into the network to obtain the strongly perturbed prediction probability map and the deeply represented features of the strongly perturbed image under deterministic conditions.

[0018] This invention achieves feature prediction two-level collaborative alignment strategy constraint and multi-dimensional update through steps 6 to 7.

[0019] In one embodiment, step 6 involves shallow prediction output level distribution alignment to construct the output level distribution consistency loss. The method is as follows: using a cross-view distribution alignment strategy, the weakly perturbed master prediction probability map output in the first forward propagation, i.e., in step 3 under the random deactivation state, is directly obtained. And the strong perturbation master prediction probability map output in the second forward propagation, i.e., step 5, under the off random inactivation state. Calculate the bidirectional Kullback-Leibler (KL) divergence between the distributions of the weakly perturbed master prediction probability map and the strongly perturbed master prediction probability map, and sum and average the bidirectional divergences to construct a bidirectional distribution alignment constraint, thereby obtaining the distribution alignment loss of the shallow prediction output level. This is used to constrain the continuity and smoothness of the network decision boundary, forcing the network's main prediction decision boundary to remain smooth and continuous under implicit randomness and strong external occlusion interference. To perform deep latent space feature-level contrastive distillation and construct feature-level contrastive distillation loss, the method is as follows: extract the three-dimensional feature map output from the deepest bottleneck layer of the encoder, compress the feature map to a single pixel or voxel size in the spatial dimension through adaptive three-dimensional average pooling operation, and flatten it into a two-dimensional prototype matrix. Obtain the first prototype matrix corresponding to the forward propagation of weak data perturbations. A gradient detach operation is performed on it to completely lock it, stopping gradient backpropagation, and a preset temperature parameter is applied. Softmax smooth activation is used as a stable teacher feature target. ; Obtain the second prototype matrix corresponding to the forward propagation of strong data perturbation. Apply the same temperature parameters Log-Softmax activation is then performed as a student feature prediction. ; Calculate the one-way KL divergence between the predicted student features and the target teacher features, construct an asymmetric knowledge distillation consistency constraint, and obtain the latent space feature-level prototype contrastive distillation loss. , This forces strongly perturbed features that are damaged by occlusion to actively align semantically with the distribution of clean, weakly perturbed features.

[0020] In one embodiment, the distribution alignment loss of the output level and the contrastive distillation loss of the feature level are vertically combined to construct a unified hierarchical architecture feature-distribution alignment loss, expressed as: in, For feature comparison distillation weights, For distribution alignment loss weights, For characteristic level comparative distillation loss, This is the distribution alignment loss for the output level.

[0021] In one embodiment, step 7, updating the network parameters, is specifically done as follows: the overall objective function is composed of the depth-supervised loss of the labeled image. Uncertainty correction for unlabeled images and multi-scale consistency loss and the distribution alignment loss of the output stage Distillation loss compared to characteristic level The four branches are constructed together, and the formula is as follows: in, The depth supervision loss for labeled images includes the depth supervision loss of the main prediction head and all auxiliary prediction heads. The supervised loss of each prediction head is a combination of standard cross-entropy loss and Dice loss, and is calculated only for labeled samples within the batch. Multi-scale consistency loss corrected for uncertainty. Its weight; Hierarchical architecture feature-distribution alignment loss The weight.

[0022] The aforementioned losses are described below: 1. Depth-supervised loss for labeled images The supervised loss is calculated only for labeled 3D medical image data in the current batch. To fully exploit the expressive power of each feature scale in the pyramid prediction architecture, this invention introduces a deep supervision mechanism, in which the main prediction head of the network and multiple auxiliary prediction heads mounted on different layers of the decoder jointly participate in supervised constraints. For the probability map output by each prediction head, the standard cross-entropy loss and multi-class Dice loss between it and the ground truth (hand-drawn by experts) are calculated independently. The final deep supervision loss is the weighted arithmetic mean of the losses of all prediction heads. This design can provide direct and smooth gradient guidance to the lower layers of the encoder and decoder in the early stages of training, effectively accelerating the convergence process of multi-scale features.

[0023] 2. Multiscale Consistency Loss with Uncertainty Correction This loss is primarily calculated on unlabeled datasets to bridge the gap between multi-scale predictions and the fused consensus pseudo-labels under weakly perturbated flow. The calculation steps are as follows: First, calculate the multi-scale probability maps output by the main prediction head and each auxiliary prediction head, and then fuse them with the consensus pseudo-labels generated using adaptive weighted fusion with information entropy. The mean squared error (MSE Loss) between the multi-scale predictions and the consensus pseudo-labels is calculated; simultaneously, the Kullback-Leibler (KL) divergence between the multi-scale predictions and the consensus pseudo-labels is calculated as the multi-scale prediction variance. This is used to dynamically evaluate the uncertainty of model predictions at the voxel level. The final consistency loss is achieved by introducing a variance negative exponential weight penalty term (i.e., The mean squared error is dynamically adjusted and regularized by adding the mean of the variance. This loss function automatically suppresses high-variance noise interference in easily confused areas such as anatomical boundaries and low-contrast regions through an uncertainty perception mechanism, thereby providing a high-purity consistency supervision signal.

[0024] 3. Shallow prediction output level distribution alignment loss This loss function aims to constrain the continuity and smoothness of the network's master prediction decision boundary. Specifically, it involves directly acquiring the weakly perturbed prediction probability map output by the master prediction head during the first forward propagation (with random inactivation enabled and a weakly perturbed image input). And the strong perturbation prediction probability map output by the master prediction head in the second forward propagation (with the random inactivation state explicitly turned off and the input Cutout strong perturbation image). Calculate the two-way KL divergence between these two probability distributions, i.e., with respect to... Calculate the pair based on the baseline The divergence, and with Calculate the pair based on the baseline The divergence is calculated, and the results of the two calculations are summed and averaged. This loss, by applying a bidirectional distribution latch across the view and across internal states, forces the network to output a continuous decision boundary with a high degree of spatial topology consistency when facing external spatial defects and internal implicit random noise.

[0025] 4. Deep Hidden Space Feature-Level Contrastive Distillation Loss This loss is calculated in the deepest bottleneck layer of the network encoder to maintain the semantic alignment of feature representations in the high-dimensional latent space. The steps are as follows: First, extract the 3D feature maps of the bottleneck layer generated by the network for weakly and strongly perturbed image inputs. Then, use adaptive average pooling to compress and flatten these maps into a one-dimensional global prototype vector to eliminate spatial bias. Next, perform gradient detach on the first prototype vector corresponding to the weakly perturbed flow to completely lock it, stopping gradient backpropagation and using it as an absolutely stable teacher target. Simultaneously, divide the target vector by a preset temperature parameter. And perform Softmax activation; divide the second prototype vector corresponding to the strong disturbance flow by the same temperature parameter. Log-Softmax activation is then applied as the student prediction. Finally, the one-way KL divergence between the student prediction distribution and the teacher target distribution is calculated. This loss forces the heavily perturbed features, severely damaged by Cutout, to actively align semantically with clean, weakly perturbed features, fundamentally solving the problem of high-dimensional feature representation shift and semantic collapse in unlabeled data under extreme conditions.

[0026] This step uses the backpropagation algorithm to update the parameters of the entire network. Specifically, the AdamW optimizer is used to perform gradient backpropagation to minimize the total objective function, and the learning rate of the network is dynamically adjusted using a polynomial decay strategy throughout the training lifecycle.

[0027] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention addresses the representation shift problem of deep features under strong perturbations: By introducing an asymmetric teacher-student feature contrast distillation mechanism, this invention utilizes features extracted with weak perturbations as stable anchor targets, forcing semantic alignment between strongly perturbated features and these targets. Even if the input image is severely occluded, the purity of the feature distribution in the high-dimensional latent space can be guaranteed, effectively preventing feature collapse of unlabeled data.

[0028] 2. Breaking the vicious cycle of false label confirmation bias: This invention combines pyramid multi-scale prediction with information entropy-based adaptive fusion to form an uncertainty perception system. The information entropy weighting mechanism naturally suppresses high uncertainty predictions at fuzzy boundary conditions, and combined with the multi-scale prediction variance penalty term, it filters out erroneous supervision signals at the source, significantly improving the quality of consensus false labels.

[0029] 3. Enhanced robustness and decision smoothness in dealing with clinical artifacts: By applying mask-based strong occlusion (Cutout), the network is forced to break away from local texture dependencies and learn the global topology; at the same time, combined with cross-view bidirectional distribution alignment constraints, the model is forced to maintain consistent output under conditions of closed implicit randomness and strong external disturbances, effectively eliminating oscillations in the decision boundary.

[0030] 4. Superior performance with minimal annotations: Experimental results show that in extreme scenarios with only a small amount of labeled data (e.g., 10% or less), the method of this invention significantly outperforms existing advanced semi-supervised segmentation methods in core evaluation metrics such as the Dice coefficient and 95% Hausdorff distance on the ACDC heart segmentation and BraTS2019 glioma segmentation tasks when dealing with medical image segmentation tasks with complex shapes and blurred edges. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0032] Figure 1 This is a schematic diagram of the overall process of the semi-supervised image segmentation method proposed in a preferred embodiment of the present invention, showing the main stages from data preparation to model inference.

[0033] Figure 2 This is a detailed structural diagram of the semi-supervised medical image segmentation method framework based on two-level collaborative alignment of uncertainty perception and feature prediction proposed in a preferred embodiment of the present invention, which clarifies the interaction relationship between data flow, feature flow and network.

[0034] Figure 3 This is the part that verifies the implementation effect of the present invention. It is a visual comparison chart of the segmentation results of the present invention method and a variety of existing advanced semi-supervised methods on a 10% labeled ACDC cardiac MRI dataset, presented in 2D slice form.

[0035] Figure 4 This section verifies the effectiveness of the invention. It presents a visual comparison of the segmentation results of the method of the present invention with those of various existing advanced semi-supervised methods on the 10% labeled BraTS2019 glioma segmentation dataset, using 3D slicing and surface rendering. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the present invention clearer and more complete, the technical solutions of the present invention will be further described in a non-limiting manner below with reference to the accompanying drawings and specific embodiments.

[0037] Preferred embodiment of the invention: This invention proposes a semi-supervised medical image segmentation method based on two-level collaborative alignment of uncertainty perception and feature prediction. This method is specifically designed for handling complex medical anatomical structures, possessing strong anti-interference and feature extraction capabilities, and is particularly suitable for high-precision segmentation. The implementation process of this invention is described in detail below with specific steps: like Figure 1 As shown in this embodiment, a semi-supervised medical image segmentation method based on dual-level collaborative alignment of uncertainty perception and feature prediction is described in detail. The overall implementation process mainly includes the following six steps: Step 1, data preparation and network initialization, including image resampling and normalization, and constructing a single-network pyramid segmentation architecture containing multi-scale auxiliary prediction heads; Step 2, constructing strong and weak perturbation dual views, preserving standard data for unlabeled images to enhance the generation of weak perturbation streams, and applying strong perturbation streams based on pixel or voxel mask occlusion; Step 3, performing adaptive fusion based on information entropy, outputting a multi-scale probability map from the network, and calculating pixel or voxel... Step 4: Assign weights based on information entropy, aggregate to generate consensus pseudo-labels, and construct consistency correction constraints; Step 5: Implement cross-view bidirectional distribution alignment, explicitly disable the random inactivation module within the network to extract strong perturbation predictions, and calculate the output-level bidirectional KL divergence to smooth the decision boundary; Step 6: Perform asymmetric feature-level contrastive distillation, use gradient truncation to lock weakly perturbation deep features as teacher anchors, force strong perturbation features to perform semantic alignment to them, and jointly optimize the overall objective function with multiple losses; After iterative training until convergence, Step 7: Validate the implementation effect on an independent test set, and evaluate metrics such as Dice coefficient and HD95 to verify model performance. Detailed process description is as follows: Step 1: Data Preparation and Preprocessing (1) Dataset acquisition and partitioning To verify the effectiveness of the method on data of different dimensions, this embodiment selected the internationally publicly available two-dimensional ACDC dataset and the three-dimensional BraTS2019 dataset as experimental data sources. ACDC (2D) contains cardiac Cine-MR sequences of 100 subjects, including three anatomical labels: left ventricle (LV), right ventricle (RV), and myocardium (MYO). The dataset was divided into a training set of 70 cases (with only 3 or 7 cases as labeled data, and the rest as unlabeled data), a validation set of 10 cases, and a test set of 20 cases. The BraTS2019 (3D) dataset comes from the 2019 Tumor Segmentation Challenge (BraTS2019) and contains three-dimensional MRI images of 335 patients with brain tumors. Each data sample contains a T1-weighted sequence (which cannot clearly show the lesion but facilitates observation of the overall anatomical structure), a T1-enhanced sequence (T1ce) (further processing based on the T1 sequence to make the tumor tissue edge more obvious), a T2-weighted sequence (which can clearly distinguish the tumor core from the surrounding edema area, but the high signal of cerebrospinal fluid may interfere with edge determination), and a fluid attenuation inversion recovery sequence (FLAIR) (effectively highlighting the boundary between the edema area around the tumor and the non-enhanced tumor area). Furthermore, each voxel in this dataset may correspond to four different annotations: the edema area around the brain tumor, the enhanced tumor, the non-enhanced tumor, and the image background. Therefore, the model needs to annotate these four types separately.

[0038] (2) Standardized pretreatment To eliminate data heterogeneity and adapt to network input, the following preprocessing is performed on all data: 1. Resampling and Normalization: Resample all data to a fixed resolution. Truncate the voxel intensity values ​​(e.g., 0.5%-99.5% quantile) and normalize them to the interval [0, 1] or [-1, 1].

[0039] 2. Spatial Cropping and Extraction: For 2D tasks, the 3D data is divided into independent 2D slices along the minor axis and uniformly cropped or filled to a fixed size. For 3D tasks, during training, fixed-size 3D blocks are randomly cropped from the original 3D volume as network input to adapt to memory limitations.

[0040] 3. Online data augmentation: Build a data generator to apply augmentation operations such as random rotation and random scaling in real time to expand the diversity of training data.

[0041] Step 2: Constructing a multi-scale pyramid segmentation network and its internal state control mechanism like Figure 2As shown, this invention constructs a multi-scale pyramid segmentation model based on single-network double forward propagation collaboration.

[0042] (1) Network infrastructure The medical image segmentation network is a single-network model. For 2D tasks such as ACDC, the U-Net architecture is preferred; for 3D tasks such as BraTS, the V-Net architecture is preferred. Both include: 1. Encoder: Used for feature extraction. It consists of successive downsampling modules that use convolutional layers in conjunction with pooling layers to progressively extract deep semantic features. As the network depth increases, the spatial resolution of the feature maps decreases, the number of channels multiplies, and a bottleneck layer containing global high-dimensional semantics is formed at the deepest point.

[0043] 2. Decoder: Used for resolution restoration. Symmetrical to the encoder, it gradually restores spatial dimensions using 3D upsampling operations and fuses shallow detail features from the corresponding encoder levels through skip connections.

[0044] 3. Pyramid Multi-Scale Prediction Head: A main prediction head is placed at the very end of the decoder, while auxiliary prediction heads are attached at different feature scale levels (e.g., at 1 / 2, 1 / 4, and 1 / 8 resolution) to map the feature maps to... One channel ( (where the number of categories is ), together they output a multi-scale voxel-level predicted probability map.

[0045] (2) Internal state control and two-way forward propagation mechanism During training, this invention does not use exponential moving average (EMA) to update the second network. Instead, it explicitly adjusts the random deactivation state variable within the network and completes two asymmetric forward propagations within the same training batch: the first forward propagation normally enables random deactivation to introduce implicit randomness to assess uncertainty; the second forward propagation forcibly disables random deactivation (setting the deactivation probability to 0), making the network a deterministic backbone model that focuses on dealing with strong external interference.

[0046] Step 3: Construct an asymmetric input stream with strong and weak perturbations In order to construct effective consistency constraints and feature alignment on unlabeled data, this invention designs an asymmetric "strong and weak perturbation" input stream construction strategy.

[0047] (1) Weak data perturbation (for the first forward propagation) Sample a batch of 3D images from an unlabeled dataset. We apply weak data perturbations to the network. These perturbations consist only of standard 3D random flipping, random rotation, and random pruning. This slight perturbation preserves the complete anatomical structure and semantic information, aiming to allow the network to generate clean, benchmark multi-scale predictions with random deactivation enabled, serving as the target for subsequent consistency learning.

[0048] (2) Strong data perturbation (for the second forward propagation) For the same batch of images A highly destructive data perturbation is applied, and the perturbed image is input into a network in a deterministic state (random inactivation disabled). This embodiment preferably employs a strong perturbation strategy based on mask occlusion (Cutout), the specific operation of which is as follows: 1. Mask Generation: For each image tensor in the batch, first initialize a mask tensor (Mask) of the same size with all 1s. Randomly generate a fixed-size 2D rectangular or 3D cubic region in 2D / 3D space, and forcibly set the mask values ​​within the region to 0.

[0049] 2. Image Occlusion Mixing: The original weakly perturbated image is multiplied element-wise with the mask tensor to generate a strongly perturbated image. Since the value of the cube region in the mask is 0, this operation artificially creates a pure black occlusion block in the original image, realistically simulating severe local tissue loss or artifact interference in clinical scans. This mask-based strong perturbation completely cuts off the model's dependence on local textures, forcing it to infer based on global contextual topological information, significantly enhancing the model's robustness.

[0050] Step 4: Fusion of Pyramid Uncertainty Perception and Consensus Pseudo-Label To address the identification bias of pseudo-labels at lesion edges and in low-contrast areas, this invention proposes a pseudo-label adaptive fusion and uncertainty correction mechanism based on multi-scale information entropy.

[0051] (1) Adaptive fusion computing based on information entropy During the initial forward propagation, the network's main prediction head and each auxiliary head output multi-scale weakly perturbated prediction probability maps. This invention does not blindly adopt a single output, but independently calculates the information entropy of each scale prediction at the voxel level. To measure uncertainty: in To prevent division by zero of small constants (such as 1) ).

[0052] Next, the information entropy is transformed into fusion weights using a negative exponential mapping function. : A higher weight indicates a more certain prediction from the model for that voxel. The probability plots at each scale are weighted and averaged to generate a high-purity consensus pseudo-label. This mechanism automatically assigns extremely high weight to clear boundaries while depriving ambiguous areas of their voice.

[0053] (2) Multiscale consistency of uncertainty correction Constraint calculation of prediction probability maps at various scales Consensus Pseudo-labels The KL divergence between the values ​​is summed to obtain the multiscale prediction variance of the voxel. Constructing a multi-scale consistency loss with uncertainty correction. : Using the negative exponent of variance As a penalty, the proportion of high uncertainty regions in the mean squared error is automatically reduced to prevent the erroneous guidance of noisy gradients.

[0054] Step 5: Implicit Elimination and Cross-View Multidimensional Consistency Alignment To address the issues of model decision boundary oscillations and feature representation shifts under strong perturbations, this invention implements the following two-dimensional cross-view alignment and joint optimization strategies.

[0055] (1) Output level: bidirectional distribution alignment across views Obtain the master prediction probability map generated by enabling random deactivation and inputting weak perturbations. And the master prediction probability plot generated by disabling random deactivation and strong input perturbations. Calculate the bidirectional KL divergence between the two to construct the distribution consistency loss. : The bidirectional constraint forcibly locks the output manifold of the two forward propagations, forcing the model to maintain an extremely smooth and continuous decision boundary under implicit randomness and strong external occlusion interference.

[0056] (2) Characteristic level: Asymmetric prototype comparative distillation The 3D feature map output from the deepest bottleneck layer of the network is extracted and flattened into a prototype matrix through adaptive 3D average pooling.

[0057] 1. Teacher Anchor Point Construction: Obtaining the First Prototype Matrix Corresponding to the Weakly Perturbed Flow Gradient truncation is applied to it, and a temperature parameter is applied. Convert to teacher probability distribution .

[0058] 2. Comparative Feature Distillation: Obtaining the Second Prototype Matrix Corresponding to Strongly Disturbed Flows Transformed into the logarithmic probability distribution of students .

[0059] Calculate one-way divergence This forces severely occluded features to actively perform deep semantic alignment with clean, weakly perturbed features.

[0060] In this embodiment, the above-mentioned output-level cross-view distribution alignment loss existing in the shallow layers of the network is used. Distillation loss compared to feature-level prototypes existing deep within the network Vertical combination is defined as hierarchical architecture feature-distribution alignment loss. This is used to jointly smooth the decision boundary and eliminate latent space representation bias.

[0061] Finally, combined with deep supervision loss on labeled data. Uncertainty Consistency Loss Hierarchical architecture feature distribution alignment loss Construct the overall objective function. Update all network parameters using the backpropagation algorithm. Repeat the above steps until the model converges after the maximum number of iterations.

[0062] Step 6: Verify the Implementation Results To fully and objectively demonstrate the effectiveness and advancement of the "semi-supervised medical image segmentation method based on dual-level collaborative alignment of uncertainty perception and feature prediction" proposed in this invention, this embodiment conducts a comprehensive performance comparison with various current advanced semi-supervised segmentation methods on a standardized experimental platform.

[0063] (1) Experimental platform and parameter settings 1. Hardware environment: All experiments were conducted on a high-performance computing platform equipped with an Intel(R) Core(TM) CPU and an NVIDIA GeForce RTX 5060 Ti GPU (16GB VRAM).

[0064] 2. Software environment: The operating system is Windows 11, the deep learning framework is PyTorch 2.8.0, and the parallel computing library is CUDA 12.8.

[0065] 3. Training parameters: For the ACDC dataset (2D), the input image size is uniformly adjusted to... The batch size was set to 24, and the maximum number of iterations was 30,000. The optimizer used was AdamW with an initial learning rate of 0.0001 and weight decay of 0.0001. For the BraTS2019 dataset (3D), the input block size was cropped to... The batch size is set to 4, the maximum number of iterations is 30,000, the optimizer is AdamW, and the initial learning rate is 0.0001. General hyperparameter settings are as follows: 1. Consistency Ramp-up: The maximum weight coefficient of the multi-scale consistency loss with uncertainty correction is set to 0.1. This weight is not fixed, but adopts a sigmoid ramp-up strategy, which makes the consistency constraint weaker in the early stage of training and increases smoothly with the number of iterations to prevent the model from fitting noise in the early stage.

[0066] 2. Distribution Alignment Weight: The weight of the cross-view bidirectional KL divergence loss is fixed at [value missing]. .

[0067] 3. Contrastive Weight: The weight of the asymmetric Teacher-Student prototype contrastive distillation loss is fixed at [value missing]. .

[0068] 4. Contrastive Temperature: In calculating the prototype comparative distillation, the temperature scaling parameter is set to adjust the smoothness of the Softmax probability distribution. .

[0069] (2) Evaluation metrics This embodiment uses four internationally recognized medical image segmentation evaluation metrics to comprehensively measure the model performance: 1. Dice Coefficient (%): A metric that measures the degree of volume overlap between the predicted segmented region and the actual region; a higher value is better.

[0070] 2. Jaccard Index (%): Also known as the Intersection over Union (IoU), it also measures the degree of volume overlap, and a higher value is better.

[0071] 3. 95% Hausdorff distance (95% HD, mm / voxel): A robustness metric that measures the maximum mismatch between the predicted and actual boundaries, excluding 5% outliers; a lower value is better. This metric accurately reflects the model's ability to characterize organ boundaries.

[0072] 4. Average Surface Distance (ASD, mm / voxel): Measures the average distance between the predicted boundary and the actual boundary; a lower value is better.

[0073] (3) Experimental results and analysis on the ACDC dataset (2D cardiac MRI): Tables 1 and 2 show the performance comparison of the method of the present invention with the baseline model and mainstream semi-supervised methods on the ACDC dataset. Table 1 Table 2 Under the extremely scarce 5% annotation setting shown in Table 1, the purely supervised baseline U-Net model suffers from severe overfitting, resulting in a performance collapse with a Dice coefficient of only 47.83%. In contrast, the method of this invention exhibits strong robustness, improving the Dice coefficient to 88.90% and the Jaccard index to 81.12%. Compared with similar state-of-the-art methods, this invention achieves superior performance. Specifically, this invention outperforms the state-of-the-art MPER (88.14%) and CPAM-TRCM (88.62%) methods by 0.76% and 0.28% respectively on the Dice metric. More importantly, on the 95% Hausdorff distance (HD95) metric, which measures the quality of boundary segmentation, this invention achieves an excellent score of 1.72, significantly lower than BCP (4.23), CrossMatch (3.15), and MPER (4.24). The significant reduction in HD95 strongly demonstrates that the "multi-scale adaptive fusion based on information entropy" and "cross-view bidirectional distribution alignment" strategies proposed in this invention can effectively suppress high uncertainty noise in pseudo-labels, making the segmentation boundary more closely match the true edge of the anatomical structure, and avoiding the boundary oscillation and artifact problems common in traditional methods.

[0074] Under the 10% annotation setting shown in Table 2, the performance of this invention is further improved, with the Dice coefficient exceeding 90.06% and the HD95 further reduced to 1.19. This result is very close to the upper limit of fully supervised U-Net (91.44%), and outperforms the comparison methods such as CrossMatch, MPER, and CPAM-TRCM in various metrics. This indicates that with a slight increase in the amount of labeled data, this invention can more efficiently mine features from unlabeled data under external interference, achieving rapid convergence of model performance.

[0075] Furthermore, Figure 3 This paper presents a visual comparison of segmentation results from different semi-supervised methods on the ACDC dataset. In basal slices with complex cardiac anatomy, due to scarce labels, contrastive methods (such as BCP and CrossMatch) easily confuse features of the right ventricle (RV) with the background, resulting in obvious artifacts and boundary adhesion around the RV wall. In contrast, the segmentation map generated by the method of this invention is more consistent with the true labels, the boundary contour of the right ventricle is smoother, and the misclassified regions far from the main body of the organ are effectively eliminated. This confirms that the "consensus pseudo-label fusion based on information entropy" strategy proposed in this invention can effectively suppress the error amplification in blurred regions, while "cutout strong perturbation" enables the network to learn a global anatomical topology resistant to interference. Contrastive methods exhibit discontinuities in the continuity of the myocardial wall, while this invention can restore the annular structure of the myocardium, indicating that the "asymmetric teacher-student feature contrast distillation" strategy significantly enhances the model's semantic alignment ability of deep latent space features under strong perturbation, greatly improving the capture and reconstruction stability of subtle anatomical structures.

[0076] (4) Experimental results and analysis on the BraTS2019 dataset (3D brain tumor MRI): Tables 3 and 4 show the performance of the method of the present invention on the 3D brain tumor (glioma) segmentation task. Since the 3D task has a large memory footprint and the training batch is usually small, it is easy to cause unstable feature alignment. Therefore, this experiment focuses on verifying the effectiveness of the "contrastive feature distillation" mechanism of the present invention.

[0077] Table 3 Table 4 Under the 10% (25 cases labeled) scarce annotation setting shown in Table 3: the purely supervised baseline V-Net model suffers from severe overfitting, resulting in a Dice coefficient of only 74.43%. In contrast, the method of this invention achieves a Dice coefficient of 86.21% and a Jaccard index of 77.85%, significantly outperforming mainstream semi-supervised methods such as ETC-Net (85.55%), SGRS-Net (85.67%), and the latest DMED-Seg (85.51%). The 95% Hausdorff distance (95HD) of this invention is reduced to 5.43, while all the comparative methods have a similar index of 8.27 or higher. Regarding the average surface distance (ASD), this invention achieves only 1.39, which is superior to the second-best method, ETC-Net's 1.84. The above data strongly supports the core viewpoint of this invention: Under complex medical images and small-batch training conditions, traditional methods are prone to representation shifts in deep features due to local artifact interference or implicit randomness; while the feature-level contrastive distillation strategy introduced in this invention, combined with cross-view bidirectional distribution alignment, forces feature alignment by locking pure teacher anchor points, which effectively maintains the global stability of feature representation and significantly reduces the breakage and adhesion of predicted structures.

[0078] Under the 20% (50 cases labeled) annotation setting shown in Table 4: the performance of this invention further improved, with the Dice coefficient reaching a strong 87.52% and the Jaccard coefficient increasing to 78.41%. Furthermore, this invention achieved scores of 5.10 and 1.27 on the two core edge quality metrics, 95HD and ASD, respectively. These results fully demonstrate that the semi-supervised regularization mechanism of this invention not only effectively utilizes unlabeled data but also surpasses simple strongly supervised fitting in noise-resistant manifold learning, enabling the model to obtain smoother and more accurate anatomical boundaries than the fully supervised model.

[0079] Figure 4Further, 3D surface renderings and cross-sectional images demonstrate the segmentation quality of various methods on the BraTS 2019 brain tumor dataset. In experiments with a low annotation ratio of 10%, due to the highly irregular shape and extremely blurred boundaries of brain tumor lesions, the 3D surfaces generated by existing methods (such as SGRS-Net and DMED-Seg) exhibit significant topological errors, specifically structural missing points or isolated noise points at tumor edge protrusions. This is due to the limitations of feature extraction caused by traditional weak perturbations, making it difficult for the model to maintain global spatial shape constraints. In contrast, the tumor surface reconstructed by the method of this invention is smoother and denser, not only without holes in the main structure but also maintaining structural connectivity in irregular edge infiltration areas. This is thanks to the "Cutout strong perturbation" and "information entropy-based pyramid pseudo-label adaptive fusion" mechanisms introduced in this invention, which not only filter out the uncertainty noise at the lesion edges but also enable the model to learn robust global morphological features even under conditions of local defects.

[0080] In summary, this invention successfully addresses the critical issues of pseudo-label confirmation bias and feature collapse in semi-supervised medical image segmentation through an innovative multi-scale uncertainty adaptive fusion mechanism and an asymmetric cross-view feature comparison distillation strategy. It not only achieves significant performance improvements in complex medical imaging tasks but also breaks through the bottleneck of fully supervised baselines in terms of geometric boundary accuracy.

Claims

1. A semi-supervised medical image segmentation method based on two-level collaborative alignment of uncertainty perception and feature prediction, characterized in that, Includes the following steps: Step 1: Provide labeled medical image datasets and unlabeled medical image datasets; initialize a medical image segmentation network, which includes an encoder, a decoder, a main prediction head, multiple auxiliary prediction heads, and a random deactivation module; Step 2: Sample images from the dataset to construct the current batch, and apply weak and strong data perturbations of different external interference intensities to the unlabeled images in the current batch to obtain weak data perturbation images and strong data perturbation images. Step 3: Input the weak data perturbation image into the medical image segmentation network, extract deep representation features, and generate a multi-scale weak perturbation prediction probability map from the output branch; Step 4: Calculate the information entropy of each scale based on the weak perturbation prediction probability map, allocate fusion weights through information entropy and generate consensus pseudo-labels by weighted averaging; combine the prediction results of each scale with the KL divergence of the consensus pseudo-labels to estimate the uncertainty variance and construct the uncertainty-corrected multi-scale consensus loss. Step 5: Turn off the random deactivation module, input the image with strong data perturbation into the medical image segmentation network, and obtain the strong perturbation prediction probability map and the deep representation features of the strong perturbation. Step 6: Align the deep representation features of strong data perturbation with the deep representation features of weak data perturbation at the latent space feature level, align the prediction probability map of strong perturbation with the main prediction probability map of weak perturbation at the decision boundary distribution, and construct a unified hierarchical architecture feature-distribution alignment loss. Step 7: Combine the depth supervision loss of labeled images, the multi-scale consistency loss with uncertainty correction, and the hierarchical architecture feature-distribution alignment loss to construct the overall objective function, update the network parameters, and repeat steps 2 to 7 until the model converges. Output the trained network for medical image segmentation.

2. The method according to claim 1, characterized in that, The medical image segmentation network is based on a multi-scale pyramid segmentation network, which includes a single feature encoder for extracting multi-scale deep semantic features and a single feature decoder for resolution restoration. The main prediction head is set at the last output layer of the single feature decoder, and an auxiliary prediction head is attached to each of the different feature scale levels of the single feature decoder to jointly construct a multi-scale pyramid prediction architecture. By explicitly adjusting the state control variable of the random deactivation module in the network, the random deactivation module is turned on during forward propagation of images with weak data perturbation and turned off during forward propagation of images with strong data perturbation, thus achieving double asymmetric forward propagation within the same training batch. The random deactivation module is a random deactivation layer set in the medical image segmentation network, which is used to randomly mask some features according to a preset deactivation probability when it is turned on, and to set the deactivation probability to zero when it is turned off.

3. The method according to claim 1, characterized in that, In step 2, standard data enhancement is retained for the current batch of images as a weak data perturbation image; and based on the strong perturbation strategy of mask occlusion, a preset size occlusion block is randomly generated and applied to the current batch of images to generate a corresponding strong data perturbation image. The specific operation is as follows: a full-1 mask tensor with the same size as the input three-dimensional medical image block is constructed; a center coordinate is randomly selected within the mask tensor to generate a two-dimensional pixel block or three-dimensional voxel block of a preset size, and the values ​​in the pixel block or voxel block are forcibly set to zero to obtain the target mask tensor; the unlabeled medical images of the current batch are multiplied element-wise with the target mask tensor to erase the values ​​of the corresponding regions in the images to zero, simulating real clinical lesion occlusion or artifact interference.

4. The method according to claim 1, characterized in that, In step 4, the specific route for generating consensus pseudo-labels is as follows: First, obtain the multi-scale weak perturbation prediction probability maps output by the main prediction head and each auxiliary prediction head; using an adaptive weighting function based on information entropy, independently calculate the information entropy of each prediction probability map at the pixel or voxel level. The adaptive weighting function is established through a negative exponential mapping, so that the smaller the information entropy at that scale, the closer the fusion weight assigned to that scale is to 1, and the larger the information entropy, the closer the fusion weight assigned to it is to 0; finally, perform a weighted average and normalization on the multi-scale weak perturbation prediction probability maps according to the fusion weights, and aggregate to generate high-purity consensus pseudo-labels.

5. The method according to claim 1 or 4, characterized in that, In step 4, the multi-scale consensus loss method for uncertainty correction is constructed as follows: For the probability maps output by the main prediction head and each auxiliary prediction head, the KL divergence between them and the consensus pseudo-label is calculated respectively. The divergences of all scales are summed as the multi-scale prediction variance of the current batch. An uncertainty penalty term is constructed using the negative exponent of the multi-scale prediction variance. The penalty term is used as a weight to dynamically adjust the mean square error between each scale prediction and the consensus pseudo-label, thereby automatically suppressing noise interference at low contrast and blurred dissection boundaries.

6. The method according to claim 1, characterized in that, In step 5, the specific operation of shutting down the random deactivation module is as follows: before the second forward propagation, by setting the state control variable, the deactivation probability of the random deactivation module integrated in all network layers of the medical image segmentation network is forced to be specified as zero, thereby eliminating the explicit randomness of the internal structure of the network and ensuring that the strong perturbation input stream can pass through the complete feature manifold space without loss.

7. The method according to claim 1, characterized in that, In step 6, the distribution consistency loss of the output level is constructed by adopting a cross-view distribution alignment strategy to directly obtain the weakly perturbed master prediction probability map output in step 3 under the random deactivation state and the strongly perturbed master prediction probability map output in step 5 under the random deactivation state; calculate the bidirectional KL divergence between the distributions of the weakly perturbed master prediction probability map and the strongly perturbed master prediction probability map, and sum and average the bidirectional divergences to construct a bidirectional distribution alignment constraint, which forces the network's master prediction decision boundary to remain smooth and continuous under implicit randomness and strong external occlusion interference; The method for constructing a feature-level contrastive distillation loss is as follows: extract the three-dimensional feature map output from the deepest bottleneck layer of the encoder, compress the feature map to a single pixel or voxel size through an adaptive three-dimensional average pooling operation, and flatten it into a two-dimensional prototype matrix. Obtain the first prototype matrix corresponding to the forward propagation of weak data perturbation, perform gradient truncation operation on it to stop gradient backpropagation, and apply a preset temperature parameter to perform Softmax smooth activation as a stable teacher feature target. Obtain the second prototype matrix corresponding to the forward propagation of strong data perturbation, apply the same temperature parameter and perform Log-Softmax activation, and use it as the student feature prediction. Calculate the one-way KL divergence between the predicted student features and the target teacher features, construct an asymmetric knowledge distillation consistency constraint, and force the strongly perturbed features damaged by occlusion to actively align semantically with the clean, weakly perturbed feature distribution.

8. The method according to claim 7, characterized in that, The distribution alignment loss of the output level and the contrastive distillation loss of the feature level are vertically combined to construct a unified hierarchical feature-distribution alignment loss, expressed as: in, For feature comparison distillation weights, For distribution alignment loss weights, For characteristic level comparative distillation loss, This is the distribution alignment loss for the output level.

9. The method according to claim 8, characterized in that, In step 7, the specific method for updating the network parameters is as follows: the overall objective function is constructed jointly by four branches, and its formula is: in, The depth supervision loss for labeled images includes the depth supervision loss of the main prediction head and all auxiliary prediction heads. The supervised loss of each prediction head is a combination of standard cross-entropy loss and Dice loss, and is calculated only for labeled samples within the batch. Multi-scale consistency loss corrected for uncertainty. Its weight; Hierarchical architecture feature-distribution alignment loss The weight.

10. The method according to claim 1, characterized in that, In step 1, standardization preprocessing is performed on all image data, including resampling the images to a uniform voxel spacing, truncating gray values, and performing online data augmentation; In step 2, the standard data augmentation includes random rotation and random cropping; In step 7, gradient backpropagation is performed using the AdamW optimizer to minimize the overall objective function, and the learning rate of the network is dynamically adjusted using a polynomial decay strategy throughout the training lifecycle.

Images