A medical image segmentation method based on multi-branch distillation

Abstract

The application belongs to the field of image processing, and discloses a medical image segmentation method based on multi-branch distillation, which comprises the following steps: constructing and training a multi-branch collaborative distillation image segmentation model based on uncertainty perception, establishing a network architecture comprising a common encoder, a first decoder and a second decoder, and comprising a guided decoder and a guided decoder; differentiating Dropout disturbance on the output of the common encoder, generating multiple features and performing multi-branch forward propagation, and generating virtual labels by fusion through a CTF module; the original feature map is subjected to guided branch calculation to supervise the loss and distribution alignment loss; the output of the guided model on the disturbance feature map is subjected to multi-level consistency constraint to realize multi-view structure consistency, and then the parameter weight is automatically balanced through an adaptive task balancing mechanism to construct a total loss function; a medical image to be segmented is input into the trained image segmentation model to obtain a segmentation result. The application can realize high-quality medical image segmentation under a low labeling rate.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and in particular relates to a medical image segmentation method based on multi-branch distillation. Background Technology

[0002] Medical image segmentation is a key technology in the field of medical image analysis. It involves pixel-level identification and segmentation of organs, tissues, or lesions in images such as CT and MRI, providing doctors with quantitative data for disease diagnosis, surgical planning, and efficacy evaluation. With the widespread use of medical imaging equipment, this technology has demonstrated significant application value in radiology, pathology, and many other clinical departments.

[0003] Early medical image segmentation methods primarily relied on manual delineation or traditional image processing techniques, such as thresholding, region growing, and active contour models. These methods are highly dependent on expert experience and struggle to handle complex structures and uneven grayscale levels. With the development of machine learning, data-driven methods based on graph registration and random forests have improved adaptability to different cases to some extent, but feature design still relies on prior knowledge, resulting in limited generalization ability.

[0004] In recent years, the introduction of deep learning technology has greatly promoted the advancement of medical image segmentation technology. Fully convolutional neural networks (FCNs) and their derivative models (such as U-Net and DeepLab) can automatically extract multi-level features through end-to-end learning, achieving unprecedented accuracy in many public challenges. However, these supervised methods heavily rely on large amounts of pixel-level labeled data, while labeling medical data is extremely costly and requires the participation of experienced physicians, which has become a major bottleneck restricting their development.

[0005] In semi-supervised medical image segmentation, while existing methods have alleviated the reliance on large amounts of labeled data to some extent by utilizing unlabeled data, some inherent drawbacks remain: medical images themselves are characterized by high uncertainty, blurred inter-class boundaries, and complex structures. Therefore, two key challenges in semi-supervised medical image segmentation are: first, the large number of highly uncertain regions with blurred boundaries and complex structures in medical images require special attention; second, the quality of virtual labels in unlabeled data varies greatly, and how to break free from the constraints of the traditional one-way teacher-student framework to fully unleash the rich information contained in unlabeled data is crucial. A core challenge is how to fully utilize limited labeled data and safely and efficiently extract knowledge from large amounts of unlabeled data. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a medical image segmentation method based on multi-branch distillation, achieving high-quality medical image segmentation with low labeling rates.

[0007] To achieve the above objectives, the technical solution adopted by this invention is: a medical image segmentation method based on multi-branch distillation, comprising the following steps:

[0008] S101, the medical image dataset is divided into labeled dataset and unlabeled dataset; after preprocessing the unlabeled samples in the unlabeled dataset, local spatial structure perturbation is applied; labeled samples are randomly selected from the current training batch as clean samples to provide local structural blocks; then, continuous cubic blocks are randomly sampled in the three-dimensional voxel space, and permutation, masking or modulation operations are performed on the voxels in this region to construct the first and second strongly enhanced versions of the unlabeled samples, respectively.

[0009] S102, establish a network architecture including a common encoder, a first decoder and a second decoder; the first decoder serves as an average guidance model, updating parameters through exponential moving average; the second decoder serves as the guided model for end-to-end training; differential Dropout perturbation is applied to the output of the common encoder to generate heterogeneous feature views and perform multi-branch forward propagation;

[0010] S103: Based on the prediction results obtained from the multi-branch forward propagation, pixel-level probability statistics are aggregated for each category to obtain class centers and intra-class dispersion, which are used to characterize the global category prior; at the same time, virtual labels are generated based on the covariance structure in the analysis window of the covariance fusion module.

[0011] S104: Calculate the composite supervision loss for labeled data. For unlabeled data, calculate the distribution alignment loss based on the global category prior and the output of the guided model. This loss is used to constrain the predicted distribution to be consistent with the potential distribution. The output of the guided model to heterogeneous feature views achieves internal multi-view structure consistency loss through multi-level consistency constraints. Then, the loss weights are dynamically adjusted through an adaptive task balancing mechanism to construct the overall optimization objective.

[0012] S105: Based on the total loss function, the gradient backpropagation algorithm is used to perform the optimization and update process of model parameters.

[0013] Furthermore, data augmentation of labeled data includes random rotation, flipping, and cropping to generate diverse training samples, and the data is then fed into the network in batches for training.

[0014] Weak and strong enhancement processes were applied to the unlabeled data, respectively.

[0015] Weak enhancement includes basic operations such as scale scaling, spatial geometric transformation, and intensity perturbation;

[0016] Strong enhancement introduces a three-dimensional structured block perturbation mechanism on the basis of weak enhancement to construct training samples. According to the preset voxel ratio, a continuous three-dimensional neighborhood is randomly selected in the voxel space of the unlabeled image. The three-dimensional neighborhood covers the local anatomical structure along the horizontal plane and forms a cuboid-shaped voxel block across multiple adjacent slices in the vertical direction. Subsequently, an labeled image is randomly selected from the current training batch, and the voxel block at the corresponding position of the image is embedded into the three-dimensional neighborhood after being processed by Gaussian noise superposition or intensity scaling, so as to realize the replacement of the local anatomical structure.

[0017] Furthermore, establishing the network architecture includes a common encoder, a first decoder, and a second decoder, comprising:

[0018] Both encoders use a V-Net structure to extract features, and both decoders share encoder features but have independent parameters; differential Dropout perturbation is implemented on the encoder output features to generate multiple feature views.

[0019] Furthermore, for the same input image, after the encoder generates a common feature map, it applies three different intensities of Dropout operation: no perturbation, Dropout rate 0%; weak perturbation, Dropout rate 15%; and strong perturbation, Dropout rate 55%. This generates three sets of feature views with different perturbation intensities, which are then input into the decoders of the guiding and induced models to produce diverse prediction results. Based on the multi-view prediction results, the class center and intra-class dispersion of each foreground category are calculated to characterize the intra-class probability distribution of the foreground and to perform global prior perception.

[0020] Furthermore, multi-branch forward propagation is performed, and the uncertainty of teacher and student predictions is quantified through local statistical property analysis based on the covariance fusion module CTF, including:

[0021] Calculate the covariance matrix of the guidance model and the induced model for the prediction of the unperturbed feature. Within a 2×2 local window for each pixel, calculate the average of the guidance and induced prediction probabilities respectively. and The variance of the guiding forecast, the variance of the guided forecast, and the covariance between the guiding and guided forecasts are: ;

[0022] The covariance matrix is ​​used within the same window to characterize the cooperative variation properties of both:

[0023] ;

[0024] The region weights are obtained from the covariance matrix. :

[0025] ;

[0026] Where i and j are the pixel coordinates, and det is a matrix. The determinant of;

[0027] For the pixel set Ω of category c, the predicted probability of all pixels under different perturbation views is calculated, and the probability prototype distribution of each category is constructed for global prior perception. The category prototype mean is then calculated. and category prototype variance :

[0028] ;

[0029] ;

[0030] in Let represent the predicted probability of pixel i under the m-th perturbation view. This represents the average predicted probability of the pixel across M perturbed views. The total number of pixels in category c. and The statistic is cumulatively updated during training using the exponential moving average (EMA) of the current batch.

[0031] The model captures the co-variance of local windows using matrix determinant calculations; then, it combines global prior perception to measure the intra-class dispersion of the current prediction relative to its class center; finally, it considers the absolute difference between the model predictions. and These are fine-tuning coefficients, used to generate fusion coefficients. :

[0032] ;

[0033] in, To guide the model's predicted probabilities, The predicted probability of the derived model. The mean of the class prototype. For the class prototype variance, The mean of the predicted probability;

[0034] Based on dynamically adjusted confidence thresholds, guided predictions are directly used to ensure determinism in high-confidence regions, while consensus-based collaborative fusion is initiated in regions requiring optimization to obtain high-quality virtual labels. Assisted training:

[0035] ;

[0036] in, This is the threshold that increases with the number of training rounds.

[0037] Furthermore, a triple loss function is established and end-to-end joint optimization is achieved through trainable parameters. The three loss functions include: supervised segmentation loss, distribution alignment loss, and multi-view structure consistency loss. Dynamic weights are assigned to each loss function through trainable parameters, and they are jointly optimized with the network parameters during training.

[0038] Furthermore, end-to-end joint optimization is achieved by dynamically adjusting multi-task training through trainable balancing parameters. The three types of constrained tasks include: fitting constraints on labeled data, imposing statistical distribution constraints on unlabeled data to make the prediction results approach the joint distribution defined by the virtual label and the category prototype, and imposing stability constraints on the feature output under multiple intensity perturbations. Through learnable balancing parameters, dynamic weights are assigned to the three types of constrained tasks and adjusted synchronously with the network parameters during training.

[0039] Furthermore, the training process employs a multi-perturbation progressive learning mechanism, achieving collaborative training of the model by constructing hierarchical optimization paths, including:

[0040] The supervisory signal composite layer, serving as the basic optimization path, processes the labeled data and constructs a composite supervisory loss function for the labeled data. The supervised segmentation loss is calculated by combining the true label Y with cross-entropy loss and Dice loss. :

[0041] ;

[0042] in, This is the balance coefficient, set to 0.25; Cross-entropy loss; It is the Dice similarity coefficient loss;

[0043] The prototype-guided constraint layer, serving as an intermediate optimization path, processes unlabeled data and constructs a distribution-aligned loss function. Through virtual tags Furthermore, an auxiliary regularization term is introduced in conjunction with intra-class scatter to ensure the model's stability in response to input variations; prototype-guided alignment loss. for:

[0044] ;

[0045] Where i, j represent pixel positions, and N is the total number of pixels. This represents the output of the derived model for weakly perturbated features. This represents the KL divergence, automatically considering the category dimension. Controlling the weights of regularization terms, Represents an exponential function. The mean of the class prototype. For the class prototype variance, The mean of the predicted probability is denoted as . A small positive constant.

[0046] The multi-view structural consistency layer, as a deep optimization path, constructs a structural self-calibration loss function within the guided model. By calculating the Dice difference between unperturbed and multi-perturbed predictions, the self-calibration of the feature layer and the enhancement of output consistency are achieved.

[0047] ;

[0048] in, For Dice similarity coefficient loss, The output of the derived model for the unperturbed features. The output of the derived model for weak perturbation features. This is the output of the model under strong perturbation characteristics.

[0049] Furthermore, the total loss is ultimately obtained through an adaptive task balancing mechanism. :

[0050] ;

[0051] in, It is a learnable scalar; For the purpose of supervision loss The balance parameters, For consistency loss The balance parameters, Distillation loss The imbalance parameter, for The square root, for The square root, for The square root of.

[0052] The beneficial effects of adopting this technical solution are:

[0053] This invention breaks away from the traditional one-way teacher-student supervision framework and proposes a two-way collaborative learning network. In this network, weak perturbation prediction not only acts as the guided subject for self-distillation but also serves as a guide for stronger perturbation outputs, achieving a multi-level knowledge transfer chain from "no perturbation → weak perturbation → strong perturbation." This invention establishes a triple loss function: supervised segmentation loss constrains the accuracy of labeled data predictions; distribution alignment loss maintains consistency between no-perturbation and perturbation predictions; and multi-view structure consistency loss ensures the stability of model outputs under different perturbation levels. The multi-role collaborative mechanism allows a single weak perturbation prediction to assume both guiding and guided responsibilities, improving knowledge utilization efficiency. The triple loss function, combined with a homoscedasticity uncertainty weighting mechanism, achieves a dynamic balance of contributions to each loss, eliminating the need for manual parameter tuning. Through multi-level knowledge transfer between no-perturbation and perturbation predictions, the model can acquire more information from unlabeled data, achieving more stable and robust segmentation performance under semi-supervised conditions.

[0054] This invention effectively utilizes unlabeled data to generate high-quality virtual labels under low-label conditions through regional uncertainty-aware fusion and multi-role collaborative learning mechanisms. Furthermore, it enhances the model's self-distillation capability through a multi-level perturbation knowledge transfer chain, achieving performance close to that of V-Net fully supervised training using only 10% of labeled data. Attached Figure Description

[0055] Figure 1 This is a flowchart of the model construction method in an embodiment of the present invention;

[0056] Figure 2 This is a model architecture diagram in an embodiment of the present invention;

[0057] Figure 3 This is a comparison diagram of the segmentation results in an embodiment of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described below with reference to the accompanying drawings.

[0059] In this embodiment, as Figure 1 As shown, this invention proposes a medical image segmentation method based on multi-branch distillation, comprising the following steps:

[0060] S101, the medical image dataset is divided into labeled dataset and unlabeled dataset; after preprocessing the unlabeled samples in the unlabeled dataset, local spatial structure perturbation is applied; labeled samples are randomly selected from the current training batch as clean samples to provide local structural blocks; then, continuous cubic blocks are randomly sampled in the three-dimensional voxel space, and permutation, masking or modulation operations are performed on the voxels in this region to construct the first and second strongly enhanced versions of the unlabeled samples, respectively.

[0061] S102, establish a network architecture including a common encoder, a first decoder and a second decoder; the first decoder serves as an average guidance model, updating parameters through exponential moving average; the second decoder serves as the guided model for end-to-end training; differential Dropout perturbation is applied to the output of the common encoder to generate heterogeneous feature views and perform multi-branch forward propagation.

[0062] S103: Based on the prediction results obtained from the multi-branch forward propagation, pixel-level probability statistics are aggregated for each category to obtain class centers and intra-class dispersion, which are used to characterize the global category prior. Simultaneously, virtual labels are generated based on the covariance structure within the analysis window of the covariance fusion module.

[0063] S104: Calculate the composite supervision loss for labeled data. For unlabeled data, calculate the distribution alignment loss based on the global category prior and the output of the guided model. This loss is used to constrain the predicted distribution to be consistent with the potential distribution. The output of the guided model to heterogeneous feature views achieves internal multi-view structure consistency loss through multi-level consistency constraints. Then, the loss weights are dynamically adjusted through an adaptive task balancing mechanism to construct the overall optimization objective.

[0064] S105: Based on the total loss function, the gradient backpropagation algorithm is used to perform the optimization and update process of model parameters.

[0065] Gradient calculation and backpropagation: The gradient of the parameters of the derived model is calculated through the backpropagation algorithm, including all trainable parameters of the encoder, the derived decoder, and the covariance fusion module CTF.

[0066] The parameters of the guided and subordinate models are updated using an optimizer (AdamW / SGD) based on the calculated gradients. The learning rate is dynamically adjusted according to a preset decay strategy. The subordinate model parameters are updated using an exponential moving average strategy, as shown in the formula:

[0067] ;

[0068] in The smoothing coefficients are used to ensure the stability of the guiding model parameters.

[0069] Iterative training and convergence judgment: Repeat steps S101 to S105 until the model's performance on the validation set reaches a stable state or the preset maximum number of iterations is reached, and save the optimal model for the inference stage.

[0070] As an optimization of the above embodiments, data augmentation of labeled data includes random rotation, flipping, and cropping to generate diverse training samples, and batch inputting the data into the network for training;

[0071] Weak and strong enhancement processes were applied to the unlabeled data, respectively.

[0072] Weak enhancement includes basic operations such as scale adjustment, spatial geometric transformation, and intensity perturbation.

[0073] Strong enhancement, building upon weak enhancement, introduces a 3D structured block perturbation mechanism to construct training samples. Based on a preset voxel ratio, a continuous 3D neighborhood is randomly selected within the voxel space of an unlabeled image. This 3D neighborhood covers local anatomical structures along the horizontal plane and forms cuboid-shaped voxel blocks across multiple adjacent slices in the vertical direction. Subsequently, an labeled image is randomly selected from the current training batch, and the voxel blocks at the corresponding locations in the image are embedded into the 3D neighborhood after Gaussian noise superposition or intensity scaling, thus replacing the local anatomical structures. This enhanced sample possesses continuity in 3D space, controllable scale, and trainable data source permutation characteristics, significantly enhancing the model's robustness to structural deformation and semantic ambiguity.

[0074] As an optimization of the above embodiments, such as Figure 2 As shown, the network architecture includes a common encoder, a first decoder, and a second decoder, comprising:

[0075] Both encoders use a V-Net structure to extract features, and both decoders share encoder features but have independent parameters; differential Dropout perturbation is implemented on the encoder output features to generate multiple feature views.

[0076] Specifically, for the same input image, after the encoder generates a common feature map, it applies three different intensities of Dropout operation: no perturbation (0% Dropout rate), weak perturbation (15% Dropout rate), and strong perturbation (55% Dropout rate). This generates three sets of feature views with different perturbation intensities, which are then input into the decoders of the guiding and induced models, respectively. Based on the multi-view prediction results, the class center and intra-class dispersion of each foreground category are calculated to characterize the intra-class probability distribution of the foreground and to perform global prior perception.

[0077] Specifically, multi-branch forward propagation is performed, and the uncertainty of teacher and student predictions is quantified through local statistical property analysis based on the covariance fusion module CTF, including:

[0078] Calculate the covariance matrix of the guidance model and the induced model for the prediction of the unperturbed feature. Within a 2×2 local window for each pixel, calculate the average of the guidance and induced prediction probabilities respectively. The variance of the guiding forecast, the variance of the guided forecast, and the covariance between the guiding and guided forecasts are: ;

[0079] The covariance matrix is ​​used within the same window to characterize the cooperative variation properties of both:

[0080] ;

[0081] The region weights are obtained from the covariance matrix. :

[0082] ;

[0083] Where i and j are the pixel coordinates, and det is a matrix. The determinant of;

[0084] For the pixel set Ω of foreground c, the confidence scores under different perturbation intensities are calculated, and the class center representations of each foreground are constructed for global prior perception. The class centers of each category are then calculated. and intra-class dispersion :

[0085] ;

[0086] ;

[0087] in, To represent pixels In the Confidence level under a perturbation view For this pixel in The mean under each perturbation view, Let c be the total number of pixels in the foreground. and The statistical values ​​from this training batch are continuously accumulated and refreshed during model optimization using an exponentially weighted moving average.

[0088] The model captures the co-variance of local windows using matrix determinant calculations; then, it combines global prior perception to measure the intra-class dispersion of the current prediction relative to its class center; finally, it considers the absolute difference between the model predictions. and These are fine-tuning coefficients, used to generate fusion coefficients. :

[0089] ;

[0090] in, To guide the model's predicted probabilities, The predicted probability of the derived model. As the class center, For intra-class dispersion, The mean of the predicted values;

[0091] Based on dynamically adjusted confidence thresholds, guided predictions are directly used to ensure determinism in high-confidence regions, while consensus-based collaborative fusion is initiated in regions requiring optimization to obtain high-quality virtual labels. Assisted training:

[0092] ;

[0093] in, This is the threshold that increases with the number of training rounds.

[0094] This invention quantifies the uncertainty of each spatial location or local region by analyzing the local covariance matrices of the prediction results of the guiding and guided models, thus achieving regional uncertainty perception fusion. This method does not rely on additional prediction information entropy or complex contextual features, and is computationally simple and efficient to implement.

[0095] As an optimization of the above embodiments, three levels of consistency constraints are implemented: First, the guiding model generates high-quality virtual labels based on perturbation-free features to provide stable supervision signals for the guided model; second, the weak perturbation predictions of the guided model maintain consistency with the predictions of the guiding model, and knowledge distillation is achieved through cross-entropy loss; finally, the strong perturbation predictions and weak perturbation predictions within the guided model are compared with the unperturbation predictions to calculate cross-entropy damage, thereby implementing consistency constraints and strengthening the robustness of the model to feature perturbations.

[0096] A triple loss function is established and end-to-end joint optimization is achieved through an adaptive weighting mechanism. The three loss functions include: supervised segmentation loss, distribution alignment loss, and multi-view structure consistency loss. An adaptive task balancing mechanism is adopted to assign dynamic weights to each loss function. Trainable parameters are used to optimize the network parameters together during training. This adaptive weight allocation mechanism can dynamically adjust the weight distribution according to the training progress and the convergence of each loss function. In the early stage of training, it relies more on supervised loss, and gradually increases the contribution of unsupervised consistency loss as training progresses, achieving a more balanced and stable optimization process.

[0097] Specifically, supervised basic loss is used for labeled data. Authentic Labels The loss function is calculated by combining cross-entropy loss and Dice loss. :

[0098] ;

[0099] in, This is the weighting coefficient, set to 0.25; For cross-entropy loss, the category dimension is automatically considered. It uses the Dice similarity coefficient loss, which automatically considers the category dimension;

[0100] The prototype-guided constraint layer, serving as an intermediate optimization path, processes unlabeled data and constructs a distribution-aligned loss function. Through virtual tags Furthermore, an auxiliary regularization term is introduced in conjunction with intra-class scatter to ensure the model's stability in response to input variations; prototype-guided alignment loss. for:

[0101] ;

[0102] Where i, j represent pixel positions, and N is the total number of pixels. This represents the output of the derived model for weakly perturbated features. This represents the KL divergence, automatically considering the category dimension. Controlling the weights of regularization terms, Represents an exponential function. The mean of the class prototype. For the class prototype variance, The mean of the predicted probability is denoted as . A small positive constant.

[0103] To establish consistency across multiple views based on the derived model, the model's predictions are constrained to maintain consistency under undisturbed, weakly perturbed, and strongly perturbed inputs. By calculating the Dice difference between undisturbed and multi-perturbed predictions, self-calibration of the feature layer and enhanced output consistency are achieved.

[0104] ;

[0105] in, The consistency distillation loss of the predicted output of the derived model under different perturbation characteristics is used. For Dice similarity coefficient loss, The output of the derived model for the unperturbed features. The output of the derived model for weak perturbation features. The loss is the output of the guided model for strongly perturbed features. Through multi-level consistency constraints, the loss enables the guided model to maintain stable output under different perturbations, improves the reliability of virtual labels and the model's efficiency in utilizing unlabeled samples, thereby enhancing the overall performance of the semi-supervised segmentation method.

[0106] The total loss is finally obtained through an adaptive task balancing mechanism. :

[0107] ;

[0108] in, It is a learnable scalar; For the purpose of supervision loss The balance parameters, For consistency loss The balance parameters, Distillation loss The imbalance parameter, for The square root, for The square root, for The square root of the value. Tasks with high losses will automatically have their weight reduced, while tasks with low losses or greater stability will have their weight increased. This is used to prevent degradation and ensure stable training.

[0109] This invention is highly adaptable and versatile. It does not require the design of complex structures for specific organs or image types and can be directly applied to different medical image segmentation tasks. It is compatible with 2D and 3D image input and is also suitable for high-quality virtual label generation in semi-supervised training scenarios.

[0110] This invention can improve the accuracy and robustness of semi-supervised medical image segmentation.

[0111] Experimental results show that, compared with traditional one-way guide-guided networks, the present invention improves the Dice coefficient by an average of about 2–4% on multiple medical image segmentation datasets (such as LA and ACDC), significantly reduces the Hausdorff distance, and the model is more stable under noisy or perturbation conditions.

[0112] Quantitative comparisons of the LA dataset were performed at 5% and 10% labeling settings. "This means the higher the better," and " "This indicates that the lower the better. V-Net (100% labeled data) achieved 91.14 and 5.75 in Dice and 95HD respectively under full supervision.

[0113] Table 1 Experimental results of the LA dataset

[0114]

[0115] Quantitative comparisons of the ACDC dataset were performed at 5% and 10% labeling settings. "This means the higher the better," and " "" indicates that lower is better. U-Net (100% labeled data) achieved Dice and 95HD scores of 91.44 and 4.30 under full supervision and monitoring. The results are provided as upper limit references but are not included in the main table.

[0116] Table 2 Experimental Results of the ACDC Dataset

[0117]

[0118] from Figure 3It can be seen that the method proposed in this invention significantly outperforms UniMatch and BCP methods in terms of segmentation accuracy and structural fidelity. In each key region (marked by blue circles in the figure): compared to the BCP method, this invention does not rely on clipping / pasting operations or virtual label forced supervision, avoiding edge loss or deformation caused by virtual label noise, thus preserving key boundaries more completely. Compared to UniMatch's unidirectional distillation, this invention employs a multi-role bidirectional collaborative learning framework, where predictions with different perturbation intensities guide and are guided by each other, enabling error correction and knowledge complementarity. This results in more complete segmentation of complex structural regions and low-contrast regions, with more natural morphological preservation. The overall structure is more stable, and the recovery of complex structures and low-contrast regions is more accurate.

[0119] Figure 3 As demonstrated in the various examples, this invention exhibits excellent performance in preserving the global structure, with no significant missing parts or abnormal protrusions in any joints or edge regions. Furthermore, it is noteworthy that the segmentation results of this invention demonstrate outstanding continuity and integrity, without any discrete blocks or fragmented predictions. In summary, this invention not only overcomes the virtual label noise problem and parameter sensitivity of BCP, but also surpasses the unidirectional distillation limitations of UniMatch, achieving adaptive segmentation of critical structures and complex regions, thus validating the method's advantages in accuracy and robustness.

[0120] The Bidirectional Copy-Paste (BCP) method achieves collaborative enhancement of labeled and unlabeled data through a "bidirectional copy-paste" approach, but its implementation still has the following limitations: First, BCP crops and pastes the foreground region of the unlabeled image onto the labeled image, and the resulting hybrid image uses virtual labels predicted from the unlabeled image as supervision signals. Since the prediction of unlabeled data itself may contain uncertainties or errors, directly using virtual labels may introduce noise, affecting the model's segmentation accuracy in key boundaries or low-contrast regions. Second, BCP's performance is sensitive to the size of the foreground crop, the pasting ratio, and the position; improper parameter selection may introduce bias or distortion, requiring additional parameter tuning costs. To address the virtual label noise and sensitivity to cropping / paste parameters in the BCP method, this invention aims to design a region-level uncertainty-aware fusion mechanism. This mechanism automatically generates pixel-level fusion weights based on the uncertainty of the guide-guided prediction and introduces trainable parameters to fine-tune the fusion strategy, achieving adaptive training for key boundaries and complex structural regions. This reduces the impact of virtual label noise, avoids manual adjustment of cropping or pasting parameters, and improves segmentation accuracy, training stability, and model adaptability.

[0121] The UniMatch method aims to learn more robust model parameters that are less sensitive to image changes by requiring the model to maintain consistency not only in the final output but also in the feature representation across multiple layers on images with varying enhancement intensities. This allows for deeper utilization of unlabeled data. Virtual labels are used to supervise the model's learning on strongly enhanced data, thereby expanding the effective training sample size. The implementation process is as follows: First, the input is enhanced. For the same unlabeled image, a weakly enhanced version (e.g., simple rotation and cropping) and a strongly enhanced version (e.g., applying complex color transformations and noise injection) are generated. These two enhanced images are then sequentially input into the same segmentation model for forward computation, resulting in two sets of mid-level feature maps from shallow to deep, along with the final segmentation prediction. On the selected mid-level feature maps, the similarity between the corresponding feature maps of the weak and strong enhancements is calculated (e.g., using cosine similarity or mean squared error as a metric). Finally, at the final output layer, the difference between the two predictions is calculated (e.g., using KL divergence). The discrepancy losses calculated from all the aforementioned layers (including multiple intermediate and output layers) are weighted and summed to form a total consistency loss. This total loss is combined with the supervision loss calculated on labeled data to adjust and optimize the model's weights. Addressing the error accumulation and insufficient potential exploitation caused by the unidirectional teacher-student framework of methods such as UniMatch, this invention aims to construct a multi-role, bidirectional collaborative learning framework. This framework breaks the traditional fixed "guide-guided" role allocation, enabling predictions with different perturbation intensities to learn from and influence each other, forming a more effective knowledge distillation and error correction mechanism, and fully releasing the learning potential of unlabeled data.

[0122] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A medical image segmentation method based on multi-branch distillation, characterized by, Including the following steps: S101: Divide the medical image dataset into labeled and unlabeled datasets; after preprocessing the unlabeled samples in the unlabeled dataset, apply local spatial structure perturbation; Randomly select labeled samples from the current training batch as clean samples to provide local structural blocks; Subsequently, continuous cubic blocks are randomly sampled in the three-dimensional voxel space, and voxels in this region are subjected to permutation, masking, or modulation operations to construct the first and second strongly enhanced versions of the unlabeled samples, respectively. S102: The network architecture consists of a common encoder, a first decoder, and a second decoder; the first decoder serves as the average guidance model, updating parameters through exponential moving average; the second decoder serves as the guided model for end-to-end training. Differentiate Dropout perturbations on the output of the common encoder to generate heterogeneous feature views and perform multi-branch forward propagation; S103: Based on the prediction results obtained from the multi-branch forward propagation, pixel-level probability statistics are aggregated for each category to obtain class centers and intra-class dispersion, which are used to characterize the global category prior; at the same time, the covariance structure within the local window is analyzed based on the covariance fusion module to generate virtual labels. Perform multi-branch forward propagation. Based on the covariance fusion module, CTF generates virtual labels through local statistical characteristic analysis, including: Calculate the covariance matrix of the guidance model and the induced model for the prediction of the unperturbed feature. Within a 2×2 local window for each pixel, calculate the average of the guidance and induced prediction probabilities respectively. and The variance of the guiding forecast, the variance of the guided forecast, and the covariance between the guiding and guided forecasts are: ; The covariance matrix is ​​used within the same window to characterize the cooperative variation properties of both: ； Obtaining regional weights W from the covariance matrix ij : ； where i and j are pixel horizontal and vertical coordinates, and det is the determinant of the matrix . The synergy of local windows is captured by the matrix determinant; then, the global prior perception measures the intra-class dispersion of the current prediction from the class center; finally, the absolute difference of model prediction is generated according to the fusion coefficient and are fine-tuning coefficients, and the fusion coefficient is generated : ； Where, p t To guide the model's predicted probabilities, p s The output of the derived model for the unperturbed features. As the class center, For intra-class dispersion, This is the mean value of the pixel across M perturbation views; Based on dynamically adjusted confidence thresholds, guided predictions are directly used to ensure determinism in high-confidence regions, while consensus-based collaborative fusion is initiated in regions requiring optimization to obtain high-quality virtual labels. Assisted training: ； in, The threshold increases with the number of training rounds; S104: Calculate the composite supervision loss for labeled data. For unlabeled data, calculate the distribution alignment loss based on the global category prior and the output of the guided model. This loss is used to constrain the predicted distribution to be consistent with the potential distribution. The output of the guided model to heterogeneous feature views achieves internal multi-view structure consistency loss through multi-level consistency constraints. Then, the loss weights are dynamically adjusted through an adaptive task balancing mechanism to construct the overall optimization objective. S105: Based on the total loss function, the gradient backpropagation algorithm is used to perform the optimization and update process of model parameters.

2. The medical image segmentation method based on multi-branch distillation according to claim 1, characterized in that, The labeled samples are first subjected to conventional spatial deformation and intensity perturbation to obtain diverse training samples, and the data is then input into the network in batches for training. Weak and strong enhancement processes were applied to the unlabeled data, respectively. Weak enhancements include scale scaling, spatial geometric transformations, and intensity perturbations; Strong enhancement introduces a three-dimensional structured block perturbation mechanism on the basis of weak enhancement to construct training samples; Based on a preset voxel ratio, a continuous three-dimensional neighborhood is randomly selected in the voxel space of the unlabeled image. The three-dimensional neighborhood covers the local anatomical structure along the horizontal plane and forms a cuboid-shaped voxel block across multiple adjacent slices in the vertical direction. Subsequently, an labeled image is randomly selected from the current training batch, and the voxel block at the corresponding position of the image is embedded into the three-dimensional neighborhood after being processed by Gaussian noise superposition or intensity scaling, thereby realizing the replacement of the local anatomical structure.

3. The medical image segmentation method based on multi-branch distillation according to claim 1, characterized in that, For the same input image, after the encoder generates a common feature map, three different intensities of Dropout operation are applied: no perturbation, Dropout rate 0%; weak perturbation, Dropout rate 15%; and strong perturbation, Dropout rate 55%. Three sets of feature views with different perturbation intensities are generated and input into the guiding model decoder and the guided model decoder, respectively. Based on the multi-view prediction results, the class center and intra-class dispersion of each foreground category are calculated to characterize the intra-class probability distribution of the foreground and to perform global prior perception.

4. The medical image segmentation method based on multi-branch distillation according to claim 1, characterized in that, For the pixel set Ω of foreground c, the confidence under different perturbation intensities is counted respectively, and the class center representation of each foreground is constructed to perform global prior perception, and the class center of the category is calculated and the intra-class dispersion : ； ； wherein, is the confidence of pixel i in the mth perturbed view, is the total number of pixels of foreground c, and The statistics of the current training batch are continuously accumulated and refreshed during the model optimization via an exponentially weighted moving average.

5. The medical image segmentation method based on multi-branch distillation according to claim 1, characterized in that, End-to-end joint optimization is achieved by dynamically adjusting multi-task training through trainable balancing parameters. The three types of constraint tasks include: fitting constraints on labeled data, imposing statistical distribution constraints on unlabeled data to make the prediction results approach the joint distribution defined by virtual labels and category prototypes, and imposing stability constraints on feature outputs under multiple intensity perturbations. Through learnable balancing parameters, dynamic weights are assigned to the three types of constraint tasks and adjusted synchronously with the network parameters during training.

6. The medical image segmentation method based on multi-branch distillation according to claim 5, characterized in that, The training process employs a multi-perturbation progressive learning mechanism, achieving collaborative model training by constructing hierarchical optimization paths, including: The supervisory signal composite layer, serving as the basic optimization path, processes the labeled data and constructs a composite supervisory loss function for the labeled data X. l The supervised segmentation loss L is calculated by combining the true label Y with cross-entropy loss and Dice loss. sup : ； wherein, is a balancing coefficient, set to 0.25; L CE is a cross-entropy loss; L Dice is a Dice similarity coefficient loss; The prototype-guided constraint layer, serving as an intermediate optimization path, processes unlabeled data and constructs the distribution-aligned loss function L. cons Through virtual tags Furthermore, an auxiliary regularization term is introduced in conjunction with intra-class scatter to ensure the model's stability in response to input variations; prototype-guided alignment loss L cons for: ； Where i and j represent pixel positions, and N is the total number of pixels. This represents the output of the derived model for weakly perturbated features. This represents the KL divergence, which automatically considers the foreground category dimension. The weights of the regularization term are controlled by exp(.), which represents the exponential function. Small positive constants; The multi-view structural consistency layer, as a deep optimization path, constructs a structural self-calibration loss function L within the guided model. KD By calculating the Dice difference between unperturbed and multi-perturbed predictions, the self-calibration of the feature layer and the enhancement of output consistency are achieved. ； Among them, L Dice For Dice similarity coefficient loss, The output of the derived model for the unperturbed features. The output of the derived model for weak perturbation features. This is the output of the model under strong perturbation characteristics.

7. The medical image segmentation method based on multi-branch distillation according to claim 6, characterized in that, Introducing a balance parameter Modeling the signal-to-noise ratio of three types of constraint tasks, regulating the contribution of multiple tasks to obtain the total loss L total : ； in, It is a learnable scalar; For supervised loss L sup The balance parameters, For consistency loss L cons The balance parameters, Distillation loss L KD The imbalance parameter, for The square root, for The square root, for The square root of.

Images