Passive domain image segmentation method based on style coding and structure prior

Abstract

The invention discloses a passive domain image segmentation method based on style coding and structure prior. The method comprises the following steps: S1, collecting and preprocessing an MRI image data set comprising a source domain data set and a target domain data set; s2, using 3D U-Net as a basic segmentation network; performing source domain supervision pre-training on the 3D U-Net model by using a source domain training set in the source domain data set to generate a pre-training weight; and S3, deploying the trained 3D U-Net model to a target domain for passive self-adaption, and constructing a double-branch momentum updating architecture comprising an EMA branch and a self-adaption branch, thereby constructing a target domain segmentation model. According to the method, on the premise of privacy protection without accessing source domain data, efficient and robust online self-adaption is carried out on the pre-trained source domain model, so that the target domain medical image with salient domain offset can be accurately segmented.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and medical image processing technology, specifically relating to a passive domain image segmentation method based on style coding and structural prior. Background Technology

[0002] Magnetic resonance imaging (MRI), with its superior soft tissue resolution and multimodal imaging capabilities, has become an indispensable tool in clinical diagnosis, surgical planning, and efficacy evaluation. Utilizing deep learning technology for precise and automated segmentation of MRI images can provide physicians with objective and reproducible quantitative data. Currently, advanced deep learning segmentation models (based on the U-Net architecture) have achieved segmentation accuracy approaching or even surpassing that of human experts on the high-quality, standardized public dataset BraTS. However, these models trained in "ideal environments" often experience a precipitous drop in performance when deployed in real, complex clinical settings. The core technical bottleneck behind this is the domain offset problem.

[0003] Domain shift primarily stems from differences in imaging equipment and protocols, data quality variations, and patient population heterogeneity. These factors collectively lead to a significant gap between the "source domain" data distribution observed during model training and the "target domain" data distribution seen in real-world applications. To address this issue, academia and industry have developed Source-Free UDA (SFUDA) techniques. It assumes that during the adaptation phase, source domain data is completely unavailable, relying solely on a pre-trained source domain model and unlabeled target domain data. Most existing SFUDA methods rely on self-training, where pseudo-labels are generated using the model's predictions in the target domain, and these pseudo-labels are then used to supervise the model's updates.

[0004] Existing passive domain unsupervised domain adaptation techniques have the following technical drawbacks:

[0005] 1. Insufficient style difference compensation: Existing methods usually perform feature alignment implicitly, lacking a mechanism to explicitly model and compensate for image style differences caused by cross-device and cross-protocol issues, resulting in the feature drift problem remaining serious.

[0006] 2. Accumulation of pseudo-label noise: When the domain offset is large, the pseudo-labels initially generated by the model are of poor quality and contain a lot of noise. If they are used directly for self-training, it will lead to "adding errors to errors," with errors accumulating continuously, eventually causing the training process to oscillate or even collapse.

[0007] 3. Lack of structural consistency: Relying solely on pixel-level pseudo-label supervision, the segmentation results predicted by the model often violate basic anatomical prior knowledge. For example, what should be a single, coherent tumor region is incorrectly segmented into multiple discrete fragments, or there are unwarranted "cavities" inside the tumor, which seriously affects the clinical usability of the segmentation results. Summary of the Invention

[0008] In view of the above-mentioned deficiencies of the prior art, the present invention provides a passive domain image segmentation method based on style coding and structural prior. Under the premise of privacy protection without accessing the source domain data, it can accurately segment target domain medical images with significant domain shifts by performing efficient and robust online adaptation on a pre-trained source domain model.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] A passive domain image segmentation method based on style coding and structural prior includes the following steps:

[0011] S1. Acquire and preprocess MRI image datasets including source domain datasets and target domain datasets;

[0012] S2. Use 3D U-Net as the basic segmentation network; use the source domain training set in the source domain dataset to perform source domain supervised pre-training on the 3D U-Net model to generate pre-trained weights;

[0013] S3. Deploy the trained 3D U-Net model to the target domain for passive adaptation, and construct a dual-branch momentum update architecture that includes EMA branch and adaptive branch, thereby constructing a target domain segmentation model.

[0014] Preferably, the preprocessing in S1 includes:

[0015] The original samples in the MRI image dataset are symmetrically filled, resampled to 128×128×128, and intensity is normalized in non-zero regions to generate three MRI modalities. The four MRI modalities of T1-weighted imaging, T1ce enhancement imaging, T2-weighted imaging, and FLAIR imaging are stitched together in the channel dimension to form a 4-channel 3D image tensor.

[0016] As a preferred option, in S2, source domain supervised pre-training of the 3D U-Net model includes:

[0017] The AdamW optimizer was used with an initial learning rate of 1e-3 and a batch size of 4. The loss function for source domain supervised pre-training was the similarity coefficient loss Dice Loss. During training, random Gaussian smoothing and contrast perturbation were applied for data augmentation. A total of 120 training epochs were conducted to obtain the weights of the pre-trained model.

[0018] Preferably, S3 includes:

[0019] S31. Initialize the network parameters of the EMA branch and the adaptive branch simultaneously using the pre-trained weights;

[0020] S32. Input the original target domain samples in the target domain dataset into the EMA branch and the adaptive branch respectively, and generate pseudo-labels and segmentation prediction results respectively through the style modulation mechanism;

[0021] S33. Calculate the total loss based on the segmentation prediction result; perform backpropagation and parameter update of the adaptive branch based on the total loss;

[0022] S34. The EMA branch absorbs the parameter changes of the adaptive branch through momentum updates;

[0023] S35. Repeat S32 to S34 to perform multiple iterations on the target domain dataset to make the adaptive branch optimization converge and generate the target domain segmentation model.

[0024] As a preferred option, the specific implementation of the style modulation mechanism includes:

[0025] The original target domain sample is input into an independent style encoder, which outputs a 128-dimensional style vector. In the decoders of the EMA branch and the adaptive branch, each upsampling module embeds a style modulation unit. The style modulation unit receives the style vector and maps the style vector to a channel-level scaling factor γ and an offset factor β.

[0026] The following affine transformation is performed on the feature map F after convolution of the upsampling module based on the scaling factor γ and the offset factor β to achieve multi-scale style compensation:

[0027] F modulated =γ*F+β.

[0028] Preferably, S32 further includes: inputting the original target domain sample into the dynamic combination data augmentation module, and augmenting the original target domain sample according to the learnable weight ω to generate the augmented sample.

[0029] Preferably, the EMA branch receives the original target domain sample to obtain teacher predictions; the teacher predictions are processed by a style modulation mechanism to generate pseudo-labels for supervising the adaptive branch;

[0030] The adaptive branch receives the enhanced samples from the same original target domain samples, passes them through a style modulation mechanism, and outputs the corresponding segmentation prediction results.

[0031] Preferably, the total loss L total The calculation is as follows:

[0032]

[0033] in:

[0034] L IH For integrity loss, it is obtained by penalizing pixels in the 3D neighborhood of each segmentation prediction result whose background probability is higher than the foreground maximum probability; L CnH For connectivity loss, based on the segmentation prediction results, all connected regions are identified, the main region with the highest confidence is found, and the weighted distance from other foreground pixels to the center of the main region is penalized; L CsH The consistency loss is obtained by calculating the Dice loss between the segmentation prediction result and the pseudo-label; L aug The data augmentation loss is calculated by taking the consistency loss between the data-augmented segmentation prediction result and the same pseudo-label, and then weighted according to the augmentation strategy weights.

[0035] Preferably, S34 includes:

[0036] In each of the adaptive branches, the parameter θ A After the update, the parameter θ of the EMA branch E Instead of updating via backpropagation, the parameter changes of the adaptive branch are absorbed by the following momentum update formula:

[0037] θ Enew =p*θE old +(1-p)θ Anew

[0038] Where, θ Eold θ represents the EMA branch parameters from the previous stage. Anew is the parameter for the adaptive branch in the current stage, and p is the momentum coefficient.

[0039] Compared with the prior art, the beneficial effects of the present invention are reflected in:

[0040] 1. Unlike traditional techniques that only perform implicit feature alignment, single-branch self-training, and lack structural priors, this invention adopts a technical solution that combines style coding and decoder multi-scale modulation, dual-branch momentum update (EMA pseudo-label branch / adaptive branch), multi-head structured loss (integrity / connectivity / consistency), and learnable dynamic combination data augmentation. This enables the model to explicitly compensate for cross-device / protocol style differences and stably suppress pseudo-label noise accumulation under completely passive domain conditions (without accessing source data). It also injects anatomical priors into the optimization objective, thereby significantly improving cross-domain segmentation accuracy and structural rationality while maintaining stable training convergence.

[0041] 2. Unlike traditional techniques that use fixed / human experience-based enhancement strategies and single consistency regularization, this invention adopts a multi-view enhancement approach that dynamically combines strategies from a policy library with learnable weights. This approach is combined with EMA steady-state pseudo-labels, consistency, and structured multi-head loss for co-optimization. This allows the model to adaptively select more effective view perturbations, improve robustness to changes in the appearance of the target domain, and further reduce structural errors such as fragmentation and holes.

[0042] 3. This invention, through explicit style encoding and modulation, effectively compensates for differences in image style under different clinical environments, significantly improving the model's segmentation performance on diverse target domain data. The dual-branch momentum update strategy successfully decouples the pseudo-label generation and exploratory learning processes, effectively suppressing error accumulation during self-training and ensuring stable convergence of the adaptive process. The multi-head structured loss function incorporates anatomical prior knowledge into the optimization objective, systematically solving common problems in segmentation results such as holes, fragmentation, and blurred boundaries, making the results more clinically usable. This invention belongs to the passive domain method, requiring no access to any source domain data throughout the adaptive process, relying solely on a pre-trained model, fully meeting strict data privacy and security regulations.

[0043] In summary, this invention provides a complete, efficient, and robust solution that successfully addresses the key technical bottlenecks in the application of deep learning models in cross-domain medical image segmentation, providing important technical support for promoting the application of artificial intelligence technology in real and complex clinical environments. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the method framework of Embodiment 1 of the present invention;

[0045] Figure 2 This is a schematic diagram of the dynamic data augmentation combinatorial optimization strategy of Embodiment 1 of the present invention. Detailed Implementation

[0046] To make the technical means, inventive features, objectives, and effects of the invention readily understandable, the invention is further described below with reference to specific illustrations. However, the invention is not limited to the embodiments described below.

[0047] It should be noted that the structures, proportions, sizes, etc., illustrated in the accompanying drawings of this specification are only used to complement the content disclosed in the specification for those skilled in the art to understand and read, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0048] This invention provides an adaptive medical image segmentation method that integrates style coding, bi-branch momentum update, and multi-head structured loss under a passive domain setting. The specific implementation steps of this invention will be detailed below using a glioma MRI image segmentation task as an example.

[0049] Example 1:

[0050] like Figure 1 The passive domain image segmentation method shown includes the following steps:

[0051] S1. Acquire and preprocess MRI (Magnetic Resonance Imaging) image datasets, including source domain datasets and target domain datasets;

[0052] S11. Dataset Selection:

[0053] Source domain dataset: The publicly available dataset from the Brain Tumor Segmentation Challenge (BraTS) 2024 was selected, containing high-quality 3D multimodal MRI images of 1080 glioma cases. 1020 cases were randomly assigned as the source domain training set, and the remaining 60 cases were used for validation.

[0054] Target domain datasets: Two datasets with significant domain shifts were selected for validation: BraTS-SSA2024 (60 cases, low-quality images) and BraTS-PED 2024 (464 cases, pediatric glioma cases). Both datasets are unlabeled.

[0055] S12. Data preprocessing: Perform a uniform process on all images: symmetric filling, resampling to 128×128×128, intensity normalization in non-zero regions, and finally stitch together the four MRI modalities of T1-weighted imaging, T1ce enhancement imaging, T2-weighted imaging and FLAIR imaging (Fluid-Attenuated Inversion Recovery) in the channel dimension to form a 4-channel 3D image tensor.

[0056] S2. Use 3D U-Net as the basic segmentation network; use the source domain training set in the source domain dataset to perform source domain supervised pre-training on the 3D U-Net model to generate pre-trained weights;

[0057] The classic 3D U-Net is used as the base segmentation network. The training process is as follows:

[0058] The 3D U-Net model was trained using standard supervised training on the source domain training set prepared in S1. The AdamW optimizer was used with an initial learning rate of 1e-3 and a batch size of 4. The Dice Loss function was employed as the loss function. Data augmentation, such as random Gaussian smoothing and contrast perturbation, was applied during training. A total of 120 training epochs were performed to obtain pre-trained weights with strong segmentation capabilities, denoted as θ. src .

[0059] S3. Deploy the trained 3D U-Net model to the target domain for passive adaptation, and construct a dual-branch momentum update architecture including EMA branch and adaptive branch to build a target domain segmentation model; EMA (Exponential Moving Average) branch and adaptive branch are existing technologies;

[0060] When the trained 3D U-Net model is deployed to the target domain, the core adaptive process of this invention is initiated. During this stage, no source domain data is accessed.

[0061] S31. Model initialization:

[0062] Using the pre-trained weights θ obtained in S1 src Simultaneously initialize the network parameters for the EMA branch and the adaptive branch.

[0063] S32, take the original target domain samples x from the target domain dataset. tgt The input EMA branch and adaptive branch are used for forward propagation to generate pseudo-labels and segmentation prediction results:

[0064] Dynamic data augmentation: such as Figure 2As shown, the augmentation library includes 8 basic data augmentation strategies and parameter ranges: Gaussian smoothing σ∈[0.5,1.5], γ contrast γ∈[0.7,1.3], brightness offset Δ∈[-0.1,0.1], Gaussian noise σ_n∈[0,0.05], elastic deformation (α,σ), anisotropic scaling s∈[0.9,1.1], small angle rotation |θ|≤10°, bias field perturbation (polynomial degree≤3), generating K views from the library by combining learnable weights ω=softmax(u). Consistency loss (taking KL as an example):

[0065]

[0066] MSE / CE can also be used; weights ω are jointly learned with network parameters.

[0067] The original target domain sample x tgt Input a dynamic combination data augmentation module. This module, based on learnable weights ω, randomly samples five sets of combination strategies for the original target domain sample x. tgt Enhancement is performed, generating 5 different enhanced samples x' tgt,i .

[0068] EMA branch forward propagation: The EMA branch receives the original target domain samples X from the target domain dataset. tgt Teachers predicted After a style modulation mechanism, the segmentation prediction result is obtained, which serves as the pseudo-label y for the supervised adaptive branch. t

[0069] Adaptive branch forward propagation: The adaptive branch receives samples X from the same original target domain. tgt K enhanced views It also undergoes a style modulation mechanism to output the corresponding segmentation prediction result.

[0070] Both share the same pseudo-label Or use soft labels Calculate the consistency loss.

[0071] The specific implementation of the style modulation mechanism:

[0072] The input sample is fed into a separate style encoder, which outputs a 128-dimensional style vector s.

[0073] In the decoders of the EMA and adaptive branches, each upsampling module embeds a style modulation unit. This unit receives a style vector s and maps it to a channel-level scaling factor γ and an offset factor β.

[0074] In the decoders of the EMA and adaptive branches, each upsampling module is followed by a convolutional layer, which yields a feature map F. An affine transformation is then performed on the convolutional feature map F: F modulated =γ*F+β, to achieve multi-scale style compensation.

[0075] S33. Calculate the total loss based on the segmentation prediction results output by the adaptive branch, and perform backpropagation of the adaptive branch based on the total loss:

[0076] Multi-head loss function calculation: At the final output of the adaptive branch, three convolutional heads are connected in parallel to calculate the three types of structured losses respectively.

[0077] Integrity loss (L IH ): For each segmentation prediction result y s,i It penalizes pixels in its 3D neighborhood whose background probability is higher than the foreground probability.

[0078] Connectivity loss (L CnH ): For the segmentation prediction result y s,i It identifies all connected regions, finds the main region with the highest confidence, and penalizes the weighted distance of other foreground pixels to the center of the main region.

[0079] Consistency loss (L) CsH ): Calculate the segmentation prediction result y s,i The pseudo-label y generated by the EMA branch t The Dice loss between them.

[0080] Data augmentation loss (L aug ): Calculate the segmentation prediction results y for the 5 augmented views. s,i With the same pseudo-label y t The consistency loss between them is calculated and weighted according to the weights of the enhancement strategies.

[0081] Total loss (L) total ):

[0082] Where, λ dice The primary task is to directly maximize the overlap between the predicted and supervisory signals (true values ​​or soft pseudo-labels); it determines the convergence direction and is set to 1.0, λ. IH Used to suppress "holes" and breaks within the foreground; shifts the segmentation from "pixel alignment" to "structural rationality," set to 0.3, λ CnH To penalize multiple discrete "islands" and encourage a single main connected component, set λ to 0.2. CsH The consistency between student predictions and EMA teacher pseudo-labels is constrained to 0.5, λ. augTo improve cross-domain robustness, the model maintains its predictions invariant to appearance / geometric perturbations; therefore, it is set to 0.5.

[0083] Backpropagation and parameter update: based on L total Perform backpropagation.

[0084] During the update process, the encoder parameters of the adaptive branch and most of the convolutional layer parameters remain frozen. Only the following are updated: the style encoder, the linear layers in the style modulation unit, the convolutional layers of the three structured loss heads, and the weights ω of the dynamic data augmentation strategy.

[0085] The AdamW optimizer was used, with an initial learning rate of 5e-4 and a batch size of 1.

[0086] S34, EMA branch momentum update:

[0087] In each adaptive branch, the parameter θ A After the update, the parameter θ of the EMA branch E Instead of updating via backpropagation, the parameter changes of the adaptive branch are smoothly absorbed through the following momentum update formula:

[0088] θ Enew =p*θ Eold +(1-p)θ Anew

[0089] Where, θ Eold θ represents the EMA branch parameters from the previous stage. Anew The parameters for the adaptive branch in the current stage are set to 0.95 in this embodiment, with the momentum coefficient p.

[0090] S35. Iteration and Convergence: Repeat steps S32 to S34, performing multiple iterations on the target domain dataset. Finally, the optimized and converged adaptive branch is the final model that can accurately segment the target domain data.

[0091] The inventiveness of this invention lies in:

[0092] First, traditional self-training often relies on implicit alignment, which can easily lead to feature drift when encountering strong style differences across devices / protocols. This invention employs an independent style encoder to extract a 128-dimensional style vector and performs channel-level affine modulation on each upsampling layer of the decoder. This proactively aligns statistics at the feature level, narrowing the "appearance gap" before entering the self-training closed loop, significantly reducing initial noise from pseudo-labels. It decouples pseudo-label generation (slow EMA updates, no augmented input) from exploratory learning (multi-view augmentation). The EMA branch smooths parameters using momentum following, continuously providing more stable pseudo-labels and avoiding divergent "error-on-error" errors. Through joint constraints of integrity, connectivity, and consistency structured losses, it repairs holes, suppresses isolated areas, and maintains global consistency with pseudo-labels, making the segmentation results more anatomically consistent and improving clinical usability. Updates are only partially performed in the target domain stage (style encoder, modulation layer, multi-head convolution, augmentation weights ω), freezing the backbone and significantly reducing overfitting and instability risks while meeting privacy compliance (no access to source data).

[0093] Secondly, the coordination mechanism and the details of the landing point are not obvious:

[0094] Style encoding / modulation is commonly used in image generation or style transfer. Embedding it into the multi-scale upsampling module of a segmentation decoder to serve passive domain adaptation during testing / deployment is not intuitive and requires redesigning feature modulation and gradient flow. The combination of EMA slow branch (no enhancement) + adaptive branch (multiple enhancement) + multi-head structured loss creates a new dynamic balance between the loss term and momentum update; however, without careful design, it is prone to oscillation or collapse. This invention achieves stable convergence through momentum gating, parameter freezing, and loss weighting, a combination and ratio that is not readily available. In the unlabeled target domain, how to effectively and unbiasedly incorporate anatomical priors into the loss design and gradient path without introducing source domain data and manual annotation is a non-textbook engineering and methodological innovation.

[0095] Finally, a unified 128-dimensional style vector is used to drive the channel-level scaling / offset affine transformation of each upsampling module, and the target domain style is explicitly compensated during the segmentation network decoding stage. This approach and path are different from common front-end normalization or image-level style transfer. The EMA branch only receives pseudo-labels from the original image and does not backpropagate, while the adaptive branch receives multi-view enhancements sampled with learnable weights. The two branches are coupled with momentum average, and in the target domain stage, only the specific parameter update strategies and recipes for style / modulation / multi-head and enhancement weights, and freezing the backbone are updated.

Claims

1. A passive domain image segmentation method based on style encoding and structure prior, characterized in that, The method comprises the following steps: S1, collecting and preprocessing MRI image data sets including source domain data sets and target domain data sets; S2, using 3D U-Net as a basic segmentation network; using the source domain training set in the source domain data set to pre-train the 3D U-Net model in the source domain supervision, and generating pre-training weights; S3, deploying the trained 3D U-Net model to the target domain for source-free adaptation, constructing a double-branch momentum updating architecture containing an EMA branch and an adaptive branch, and thus constructing a target domain segmentation model. 2.The style encoding and structure prior based passive domain image segmentation method of claim 1, wherein, The preprocessing in S1 comprises: The original samples in the MRI image data set are symmetrically filled, resampled to 128*128*128, and intensity normalized in the non-zero region to generate three kinds of MRI modalities; T1 weighted imaging, T1ce enhanced imaging, T2 weighted imaging and FLAIR imaging are spliced in the channel dimension to form a 4-channel 3D image tensor. 3.The style encoding and structure prior based unsupervised domain image segmentation method of claim 1, wherein, In S2, the source domain supervised pre-training of the 3D U-Net model comprises: Using AdamW optimizer, initial learning rate 1e-3, batch size 4; the loss function of source domain supervised pre-training uses similarity coefficient loss Dice Loss; random Gaussian smoothing and contrast disturbance are applied for data enhancement in the training process; a total of 120 rounds of training are performed to obtain the pre-training model weights.

4. The style encoding and structure prior based passive domain image segmentation method according to claim 1, characterized in that S3 Comprise: S31, using the pre-training weights to initialize the network parameters of the EMA branch and the adaptive branch at the same time; S32, inputting the original target domain samples in the target domain data set into the EMA branch and the adaptive branch respectively, generating pseudo labels and segmentation prediction results respectively through the style modulation mechanism; S33, calculating the total loss according to the segmentation prediction result; performing back propagation and parameter updating of the adaptive branch according to the total loss; S34, the EMA branch absorbs the parameter changes of the adaptive branch through momentum updating; S35, repeatedly performing S32 to S34, performing multiple iterations on the target domain data set, optimizing and converging the adaptive branch, and generating the target domain segmentation model.

5. The style encoding and structure prior based passive domain image segmentation method according to claim 4, characterized in that, The specific implementation of the style modulation mechanism comprises: inputting the original target domain samples into an independent style encoder to output a 128-dimensional style vector; in the decoder of the EMA branch and the adaptive branch, a style modulation unit is embedded in each upsampling module; the style modulation unit receives the style vector and maps the style vector into a channel-level scaling factor γ and an offset factor β; according to the scaling factor γ and the offset factor β, the feature map F after convolution of the upsampling module is subjected to affine transformation as follows to realize multi-scale style compensation: F modulated = γ * F + β.

6. The style encoding and structure prior based passive domain image segmentation method according to claim 4, characterized in that, S32 further comprises: inputting the original target domain samples into a dynamic combination data enhancement module, enhancing the original target domain samples according to the learnable weight ω, and generating enhanced samples.

7. The style encoding and structure prior based passive domain image segmentation method according to claim 6, characterized in that, The EMA branch receives the original target domain samples to obtain teacher predictions; the teacher predictions are generated through the style modulation mechanism to generate pseudo labels for supervising the adaptive branch; The adaptive branch receives the enhanced sample of the same original target domain sample, passes through a style modulation mechanism, and outputs the corresponding segmentation prediction result.

8. The style encoding and structure prior based passive domain image segmentation method according to claim 4, characterized in that, The total loss L total is calculated as follows: Wherein: L IH For completeness loss, obtain by penalizing the pixels in the three-dimensional neighborhood of each of the segmentation prediction results whose background probability is higher than the maximum foreground probability;L CnH For connectivity loss, obtain by identifying all connected regions from the segmentation prediction result, finding the main region with the highest confidence, and penalizing the weighted distance of other foreground pixels to the center of the main region;L CsH For consistency loss, obtain by calculating the Dice loss between the segmentation prediction result and the pseudo label;L aug For data augmentation loss, obtain by calculating the consistency loss between the segmentation prediction result after data augmentation and the same pseudo label, and weighting according to the augmentation strategy weight.

9. The style encoding and structure prior based passive domain image segmentation method according to claim 4, characterized in that, S34 includes: At each said adaptive branch parameter θ A After update, said EMA branch parameter θ E Without backpropagation update, said adaptive branch parameter change is absorbed by the following momentum update formula: θ Enew = p * θ Eold + (1 - p) * θ Anew where θ Eold is the EMA branch parameter of the previous stage, θ Anew is the parameter of the adaptive branch of the current stage, and p is the momentum coefficient.

Images