Two-stage network-driven heart structure unsupervised domain adaptive segmentation method

Abstract

The invention discloses a two-stage network-driven unsupervised domain adaptive segmentation method for a heart structure. According to the method, a two-stage network architecture is used. In the first stage, on the basis of an ISTN image translation network, a content guidance attention fusion module is introduced to relieve edge blurring and contour distortion of a generated image, a feature enhancement up-sampling module is designed to restrain generation of artifacts in the image reconstruction process, and meanwhile, the logic consistency of a global anatomical structure is enhanced by embedding OfficientViMBlock; generating a pseudo target domain image which is adaptive to the target domain in style and retains anatomical details of the source domain; and in the second stage, a segmentation network is constructed based on nnUNetv2, a Mean Teamer double-branch architecture and a dynamic pseudo-label strategy are fused, and boundary feature alignment loss and uncertainty are introduced to guide a region hybrid module to optimize segmentation boundary precision and pseudo-label quality, so that the segmentation precision of the model can be effectively improved through the improvement.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, and particularly relates to a two-stage network driven cardiac structure unsupervised domain adaptation segmentation method. BACKGROUND

[0002] In the clinical diagnosis and treatment process of heart disease, accurate cardiac structure segmentation plays a decisive role in diagnosis accuracy, treatment plan development and efficacy evaluation. As an important sub-task of medical image segmentation, cardiac structure segmentation needs to accurately segment the fine anatomical structures such as ventricle, atrium and myocardium, and its segmentation accuracy directly affects the scientificity of clinical diagnosis and treatment decision. However, there are multiple modalities in cardiac medical imaging, and there are significant imaging mechanism differences between different modalities, for example, MRI focuses on soft tissue contrast, and CT focuses on density difference. This difference makes it difficult to directly migrate the cardiac structure segmentation model trained on one modality to another modality for cardiac image segmentation task. This cross-modality migration problem seriously restricts the wide application of deep learning models in clinical practice of heart disease, and the unsupervised domain adaptation technology provides an effective way to solve the cross-modality cardiac structure segmentation problem. Existing domain adaptation methods mainly focus on aligning the source domain and the target domain from three angles: feature alignment-based methods, image alignment-based methods and output-level alignment-based methods. However, the cardiac structure has the characteristics of complex anatomical morphology, multiple fine structures and large structural appearance difference under different modalities. When the data distribution difference between the source domain and the target domain in the cardiac structure segmentation task is too large, it is difficult to obtain satisfactory segmentation performance only by using image-level, feature-level or output-level alignment. Therefore, most domain adaptation methods for cardiac structure segmentation are committed to using joint alignment methods to achieve domain adaptation.

[0003] However, in the current unsupervised domain adaptation segmentation of cardiac structure, although various methods have made certain progress in domain shift alleviation, there are still many problems to be solved. Current unsupervised domain adaptation methods rely on static spatial alignment strategies and fail to fully consider the dynamic evolution characteristics of feature distribution during training, which often leads to blurred image boundary definition and loss of key structural details. At the same time, the existing methods have obvious shortcomings in modeling the long-distance dependence relationship and global semantic association, and it is difficult to directly capture the cross-regional anatomical structure association. In addition, the existing unsupervised domain adaptation methods only focus on improving the overall generalization performance of the model and do not optimize the boundary features, resulting in blurred segmentation mask boundary and insufficient positioning accuracy. At the same time, some co-training methods expand the model divergence by disturbing the input or network parameters, which easily produces low-confidence and high-uncertainty pseudo-labels. These unreliable pseudo-labels will mislead the network training, leading to a decrease in segmentation performance. Therefore, it is necessary to propose an unsupervised domain adaptation segmentation method to solve the above problems.

[0004] According to the search, the following prior art is found to be compared with the present application as follows:

[0005] 1. Chinese invention patent with publication number CN119649038A provides a method and device for unsupervised domain adaptation medical image segmentation based on multi-scale features. The segmentation principle is to first construct a multi-scale feature fusion segmentation network based on a residual module, which integrates a dual attention mechanism module and a multi-scale feature fusion module, and use it as a generator. Then, two discriminators are built to align the aggregated features of the source domain and the target domain with the first discriminator, and the semantic features of the source domain and the target domain with the second discriminator. By means of adversarial learning, the feature distribution difference between the source domain and the target domain is minimized to realize domain adaptation at the feature level. Relying on the architecture of the generator plus double discriminators, combined with the cross-entropy, Dice loss and L2 regularization, a hybrid segmentation loss is formed to complete the unsupervised domain adaptation segmentation of medical images.

[0006] 2. Chinese invention patent with publication number CN118735948A provides an unsupervised domain adaptation method for medical image segmentation. The algorithm principle is to first expand the data set through data augmentation, then use an adversarial generative network combined with a cycle consistency loss to realize image style conversion from the source domain to the target domain. Then, a network structure is constructed with a shared encoder plus three functionally decoupled independent decoder branches. After training the three branches with different data, three segmentation probability maps are obtained for the unlabeled target domain images. Then, high-quality pseudo labels are generated by weighting and fusing the probability maps based on the uncertainty measure of entropy. Finally, the target domain segmentation decoder is trained with the pseudo labels as the supervision signal, and the trained decoder is used to predict new target domain images to complete medical image segmentation.

[0007] 3. Chinese invention patent with publication number CN110738663A provides a dual-domain adaptation module pyramid network and an unsupervised domain adaptation image segmentation method. The algorithm principle is to build a basic segmentation network composed of an encoder, a pyramid type pooling module and a decoder. At the same time, image-level and feature-level dual-domain adaptation modules are added. The output of the pyramid type pooling module is input to the image-level domain adaptation module, and the fused features at each level of the decoder are input to the feature-level domain adaptation module. Through the adversarial training of the two domain adaptation modules, the global difference of images and the segmentation feature difference between the source domain and the target domain are reduced respectively to complete the unsupervised domain adaptation segmentation of images.

[0008] The algorithm principle of the present application is to adopt a two-stage network driving architecture, first generate a pseudo target domain image that is suitable for style adaptation and retains the anatomical details of the source domain by introducing content-guided attention fusion, feature enhancement up-sampling and EfficientViMBlock module in the generator through the improved ISTN image translation network, then build a segmentation network based on the improved nnUNetv2, fuse the Mean Teacher double branch architecture and the dynamic pseudo label strategy, combine the boundary feature alignment loss and the uncertainty guided region mixing module, and finally train the segmentation network using the pseudo target domain image combined with the source domain label and the real target domain image to realize the image segmentation of the target domain. SUMMARY

[0009] In view of the deficiencies of the existing unsupervised domain adaptation segmentation method of heart structure, the present application provides a two-stage network driven unsupervised domain adaptation segmentation method of heart structure, which adopts a two-stage network architecture to complete the unsupervised domain adaptation segmentation task, and improves the segmentation accuracy of medical images by optimizing the image translation quality and segmentation strategy.

[0010] To achieve the above-mentioned purpose, the present application provides a two-stage network driven unsupervised domain adaptation segmentation method of heart structure, comprising the following steps:

[0011] A two-stage network driven unsupervised domain adaptation segmentation method of heart structure, the method comprising the following steps:

[0012] S1, collect source domain labeled medical image data and target domain unlabeled medical image data, and perform data preprocessing;

[0013] S2, construct an improved ISTN image translation network and train it to convert the source domain image into a pseudo target domain image;

[0014] S3, construct a segmentation network of the improved nnUNetv2, and simultaneously fuse the Mean Teacher double branch architecture and the dynamic pseudo label strategy;

[0015] S4, use the pseudo target domain image combined with the source domain real label and the target domain unlabeled image as the training data of the segmentation network, perform iterative training, and detect the convergence of the model through the validation set;

[0016] S5, finally output the trained segmentation network model.

[0017] As a further improvement of this application, the improved ISTN image translation network in step S2 includes two generators, three discriminators, and two auxiliary segmentation networks. Both generators adopt an encoder-bottleneck layer-decoder structure, with only the bottleneck layer sharing weights. A content-guided attention fusion module is introduced between the encoder and the decoder to alleviate the problems of blurred edges and contour distortion in the generated images. A feature enhancement upsampling module is added to the decoder to suppress artifacts during image reconstruction. An EfficientViMBlock module is embedded in the bottleneck layer to enhance the logical consistency of the global anatomical structure and avoid the disconnect between local features and the global context.

[0018] As a further improvement of this application, the improved segmentation network in step S3 is based on nnUNetv2, which integrates the Mean Teacher dual-branch architecture and dynamic pseudo-label strategy. Specifically, the improvement of the segmentation network is achieved by introducing boundary feature alignment loss and uncertainty-guided region mixing module to optimize the segmentation boundary accuracy and pseudo-label quality.

[0019] As a further improvement to this application, the attention fusion module receives... With decoder features The specific process for adaptive feature fusion, using the input as data, is as follows:

[0020] First, the encoder features and decoder features are added element-wise to obtain the initial fused features. ,right Perform global average pooling and combine it with a bottleneck structure of two 1x1 convolutions to generate channel attention weights. At the same time, Perform average pooling and max pooling along the channel dimension separately, concatenate the pooling results, and then generate spatial attention weights through a 7x7 convolution. ;Will and By performing broadcast summation, a coarse-grained spatial importance map is obtained. ;

[0021] Then the initial fusion features Coarse-grained spatial importance map The data is concatenated along the channel dimension, and then channel shuffling, grouped convolution, and activation processing are performed sequentially to generate a fine-grained spatial importance map consistent with the input feature dimension. ;

[0022] Finally, a detailed spatial importance map was used. As a weighting factor, the encoder features With decoder features Perform dynamic modulation and output the fused feature map. The calculation formula is:

[0023] ;

[0024] Here, ☉ represents element-wise multiplication.

[0025] As a further improvement to this application, the specific operation flow of the feature enhancement upsampling module is as follows:

[0026] First, the low-resolution feature map is received and divided into 8 sub-feature groups according to the channel dimension. A dual-branch parallel architecture is used to capture multi-scale information. The first branch uses 1×1 convolution combined with one-dimensional global average pooling in the horizontal and vertical directions to encode global information. The second branch uses 3×3 convolution to capture local spatial context.

[0027] Then, through a cross-space learning mechanism, matrix multiplication and Softmax normalization are performed on the dual-branch output to generate spatial attention weights. The attention weights are multiplied element-wise with the feature groups and then upsampled by 2 times. At the same time, depth convolution is applied independently to each feature group to suppress noise and artifacts. After the feature groups are concatenated, cross-group information interaction is achieved through channel shuffling.

[0028] Finally, feature fusion is completed by adjusting the number of channels using a 1×1 convolution.

[0029] As a further improvement to this application, the EfficientViMBlock module comprises four main components: a deep convolutional layer 1, an HSM-SSD mixer, a deep convolutional layer 2, and a feedforward network FFN. Each component employs a LayerScale mechanism for feature fusion, and the fusion formula is as follows:

[0030] ;

[0031] in, For the output of the current component, This is the output of the previous component. For learnable scaling parameters, For each component.

[0032] As a further improvement to this application, the calculation process for the boundary feature alignment loss is as follows:

[0033] First, the boundary is extracted by Laplacian convolution on the true labels in the source domain. Then, the boundary is extracted after the Teacher branch prediction results are denoised by Gaussian-Laplacian operator.

[0034] Then, the feature maps of each stage of the Student branch encoder are mapped to boundary feature maps, and the boundary dimensions are matched by interpolation;

[0035] Finally, the boundary cross-entropy loss and the boundary Dice loss are summed to calculate the segmentation loss term for labeled data and the consistency constraint term for unlabeled data, respectively.

[0036] As a further improvement to this application, the uncertainty-guided region mixing module refers to calculating the entropy value based on the student and teacher model outputs to obtain an uncertainty map. :

[0037] ;

[0038] ;

[0039] in, It is the output of the model. For class probabilities;

[0040] The image is then divided into 64 regions. The three most uncertain regions from the student model's perspective are replaced with the three most certain regions from the teacher model's perspective to generate new samples. The quality of the pseudo-labels is optimized, and the generated new samples are input into the Student branch for training to enhance the learning of high-confidence features.

[0041] Beneficial effects:

[0042] Compared with existing technologies, this application effectively solves the core problems of domain differences, boundary ambiguity, and pseudo-label noise in cross-modal cardiac structure segmentation through a two-stage unsupervised domain adaptation framework. The first stage, an improved ISTN image translation network, generates high-quality pseudo-target domain images that combine the style of the target domain with the integrity of the source domain's anatomical structure through the synergistic effect of a content-guided attention fusion module, a feature enhancement upsampling module, and an EfficientViMBlock module. The second stage, an improved segmentation network, significantly improves segmentation boundary accuracy and pseudo-label quality through boundary feature alignment loss and an uncertainty-guided region mixing module, effectively enhancing segmentation accuracy and providing efficient and feasible technical support for precision medical auxiliary diagnosis. Attached Figure Description

[0043] Figure 1 This is a flowchart of the two-stage network-driven unsupervised domain adaptive segmentation method for cardiac structures in this application.

[0044] Figure 2 A comparison image showing the generator of the ISTN image translation network in this application before and after improvement;

[0045] Figure 3 This is a schematic diagram illustrating the structure of the attention fusion module, which guides the understanding of the content of this application.

[0046] Figure 4 This is a schematic diagram of the feature enhancement upsampling module structure in this application;

[0047] Figure 5 This is a schematic diagram of the EfficientViMBlock structure of this application;

[0048] Figure 6 A comparison of the MRI-CT domain adaptation performance of the model before and after improvement on the MMWHS2017 dataset. Detailed Implementation

[0049] The present application will now be described in further detail with reference to the accompanying drawings and specific embodiments:

[0050] Example 1

[0051] Reference Figure 1 This is the first embodiment of the present application, using the MMWHS2017 Heart Segmentation Public Dataset as an example. Since this dataset is already labeled, the labeling steps are omitted. This example provides a two-stage network-driven unsupervised domain-adaptive segmentation method for heart structures, with the specific steps as follows:

[0052] S1. Data preparation and preprocessing.

[0053] The MMWHS2017 dataset was divided into source domain MRI and target domain CT. The source domain contained 20 labeled MRI data points, and the target domain contained 20 unlabeled CT data points. All data were preprocessed, cropped centered on the cardiac region, and four cardiac substructures—the ascending aorta (AA), left atrial cavity (LAC), left ventricular cavity (LVC), and left ventricular myocardium (MYO)—were selected for segmentation based on their visibility in 2D views. For each 3D cropped image, the top 2% of its grayscale histogram was truncated to reduce artifact interference. Each 3D image was then normalized to a mean of 0 and a standard deviation of 1. 2D coronal slices were sampled, and data augmentation was performed simultaneously.

[0054] S2. Input the training set into the improved image translation network for iterative training.

[0055] Specifically, refer to Figures 2-5The improved ISTN image translation network mainly includes: optimizing the feature information extraction part, improving the feature upsampling operation, and efficiently modeling global structural dependencies. The optimized feature information extraction part refers to the effective fusion of low-level and high-level features. Low-level features such as edges and contours play an indispensable role in image reconstruction, but gradually lose their influence after passing through many intermediate layers. Feature fusion can enhance the information flow from shallow to deep layers, thereby preserving key anatomical structures and lesion information. A feature enhancement upsampling module is designed to replace the upsampling operation in the generator, achieving high-quality feature upsampling, reducing image artifacts and uneven pixel weight distribution, thereby improving the clarity of the generated image. At the generator bottleneck layer, EfficientViMBlock is embedded to efficiently model global structural dependencies, ensuring the global anatomical structural logic of the generated image and the spatial correlation between lesions and normal tissues.

[0056] S2.1 The optimized feature information extraction part refers to introducing a content-guided attention fusion module into the generator, adding a skip connection between the encoder and decoder, and adding the encoder features and decoder features element-wise to obtain the initial fused features. ,right Perform global average pooling and combine it with a bottleneck structure of two 1x1 convolutions to generate channel attention weights. At the same time, Perform average pooling and max pooling along the channel dimension separately, concatenate the pooling results, and then generate spatial attention weights through a 7x7 convolution. ;Will and By performing broadcast summation, a coarse-grained spatial importance map is obtained. Then the initial fusion features Coarse-grained spatial importance map The data is concatenated along the channel dimension, and then channel shuffling, grouped convolution, and activation processing are performed sequentially to generate a fine-grained spatial importance map consistent with the input feature dimension. Finally, a detailed spatial importance map was used. As a weighting factor, the encoder features With decoder features Perform dynamic modulation and output the fused feature map. The calculation formula is:

[0057] ;

[0058] Here, ☉ represents element-wise multiplication.

[0059] When W approaches 1, shallow details of the encoder are preserved; when W approaches 0, deep semantics of the decoder are preserved, achieving a content-driven adaptive balance.

[0060] S2.2 The improved feature upsampling operation refers to replacing the upsampling operation in the decoder of the generator with a feature enhancement upsampling module. This module first divides the input features into 8 sub-feature groups according to the channel dimension, and adopts a dual-branch parallel architecture. The first branch uses 1×1 convolution combined with one-dimensional global average pooling in the horizontal and vertical directions to encode global information. The second branch uses 3×3 convolution to capture local context. Through a cross-space learning mechanism, matrix multiplication and Softmax normalization are performed on the output of the two branches to generate spatial attention weights. Then, the attention weights are multiplied element-wise with the feature groups and upsampled by 2 times. Depth convolution is applied independently to each feature group to suppress noise and artifacts. After the feature groups are concatenated, cross-group information interaction is achieved through channel shuffling. Finally, the number of channels is adjusted through 1×1 convolution.

[0061] S2.3, the efficient modeling of global structural dependencies refers to embedding the EfficientViMBlock module in the bottleneck layer of the generator. This module uses a residual structure connection and contains four main components: a deep convolutional layer 1, a hidden state mixer HSM-SSD, a deep convolutional layer 2, and a feedforward network FFN. The connection of each component uses the LayerScale mechanism for feature fusion, and the fusion formula is as follows:

[0062] ;

[0063] in, For the output of the current component, This is the output of the previous component. For learnable scaling parameters, For each component.

[0064] S2.3.1 The HSM-SSD component models global long-distance dependencies with linear computational complexity. The specific process is as follows: First, linear projection is performed on the input feature X to generate the parameter matrices B and C and importance weights A required for the state-space model; then, a non-causal state-space duality mechanism is adopted to globally share the hidden states through a weighted aggregation formula. :

[0065] ;

[0066] Here, ☉ represents element-wise multiplication. The flattened input sequence is compressed into a compact hidden state sequence, achieving efficient aggregation of global information while significantly reducing computational costs. Then, a mixer containing two linear projection layers and a gating mechanism is used to perform a channel-dimensional nonlinear transformation on the hidden state. First, the hidden state is projected into content features h and gating features z, respectively. Then, element-wise multiplication and activation functions are applied. Gating modulation is performed to obtain the processed hidden state. :

[0067] ;

[0068] in, It is a linear projection layer used to perform a weighted linear transformation on the input features.

[0069] Finally, the processed hidden state is mapped back to the original input sequence space through the output projection matrix, and the features after global dependency modeling are output.

[0070] S2.3.2 The feedforward network FFN consists of two 1×1 point convolutions and a non-linear activation layer. The first point convolution expands the channel dimension of the input features to four times the original dimension. Then, the GELU activation function is used to introduce a non-linear transformation to enhance the complex feature interactions between channels. The second 1×1 point convolution compresses the channel dimension back to the original size, completing the feature dimension restoration, refining the local semantic information of the anatomical structure, and making up for the shortcomings of global modeling in the expression of local fine features.

[0071] S3. Input the pseudo-target domain image, the source domain real label, and the target domain unlabeled image into the improved segmentation network for training.

[0072] Specifically, the improved segmentation network is based on nnUNetv2, integrating a Mean Teacher dual-branch architecture and a dynamic pseudo-label strategy. The improvement of the segmentation network is specifically achieved by introducing a boundary feature alignment loss and an uncertainty-guided region mixing module to optimize the segmentation boundary accuracy and pseudo-label quality.

[0073] S3.1 The introduced boundary feature alignment loss is achieved by having the encoder learn a general boundary feature representation for labeled (pseudo-target domain and source domain labels) and unlabeled (real target domain) data. The calculation process of this loss is as follows: First, the boundary is extracted by using Laplacian convolution on the real labels in the source domain. The boundary is extracted after the prediction results of the Teacher branch of the unlabeled data are denoised by the Gaussian-Laplacian operator. Then, the feature maps of the four downsampling stages in the Student branch encoder are mapped to boundary feature maps. The boundary size is matched by interpolation. Finally, the boundary cross-entropy loss and the boundary Dice loss are added together and used to calculate the segmentation loss term for labeled data and the consistency constraint term for unlabeled data, respectively, thereby effectively suppressing the noise transmitted by the jump connections of the U-shaped network.

[0074] S3.2, The uncertainty-guided region hybrid module refers to obtaining an uncertainty map based on the calculated entropy value output by the student and teacher models. :

[0075] ;

[0076] ;

[0077] in, It is the output of the model. For class probabilities;

[0078] The image is then divided into 64 regions. The three most uncertain regions from the student model's perspective are selected and replaced with the three most certain regions from the teacher model's perspective. New samples are generated and input into the student branch for training. This allows the student network to learn more reliable feature representations from the teacher network's high-confidence predictions, while avoiding noise interference from low-confidence pseudo-labels.

[0079] Example 2

[0080] Referring to Tables 1 and 2, an embodiment of this application provides a medical image domain adaptation method based on deep learning. To verify its beneficial effects, a scientific demonstration is conducted through specific implementation methods and their effects.

[0081] The specific details of this embodiment are as follows:

[0082] The operating system is Kylin Linux Advanced Server V10, Python is 3.12, PyTorch is 2.4, the CPU is 2 Intel Xeon Gold 6326 CPUs @ 2.90GHz, the GPU is 4 NVIDIA L40 GPUs with 48GB of VRAM, and CUDA version 12.2. Detailed training configuration includes: In the first stage, training for 500 epochs with a batch size of 4. The generator and discriminator used the Adam optimizer with a learning rate of 0.0002, and the two segmenters used the SGD optimizer with a learning rate of 0.1. In the second stage, training for 200 epochs with a batch size of 32, using the SGD optimizer with an initial learning rate of 0.01, an adaptive decay strategy, and a boundary feature alignment loss weight of 1.0. The model iteratively converges to obtain the optimal model.

[0083] As described in the above embodiments, those skilled in the art will understand that this application provides a medical image domain adaptation method, and experimental verification was conducted using the MMWHS2017 dataset and an improved network model. The segmentation effects before and after the improvement are as follows: Figure 6As shown, a detection comparison experiment was conducted using the test set. It can be observed that the improved model achieves improved segmentation accuracy and demonstrates stronger robustness. Ablation experiments were performed on the first and second stages of the model, and the results are shown in Tables 1 and 2. (The baseline network in Table 1 is the ISTN image translation network plus the original nnUNetv2; the baseline network in Table 2 is the improved ISTN image translation network plus the original nnUNetv2 fused with a Mean Teacher dual-branch architecture and a dynamic pseudo-label strategy.) A comparison of the overall network before and after improvement is shown in Table 3.

[0084] Table 1. MRI-CT Stage 1 Ablation Experiment Data from the MMWHS2017 Dataset

[0085]

[0086] Table 2. MRI-CT second-stage ablation experimental data from the MMWHS2017 dataset.

[0087]

[0088] Table 3 Comparison of DICE (%) scores for each category in the MRI-CT dataset of the MMWHS2017 dataset.

[0089]

[0090] Experimental results show that, compared with the original model, the improved model performs better on the core evaluation metrics of the segmentation task: the DICE score is improved by 8.12%, and the ASD score is reduced by 27.70%. Furthermore, the model shows varying degrees of improvement in segmentation accuracy across all subcategories. These results confirm that the improvement measures can effectively enhance the model's segmentation accuracy and robustness.

[0091] The above description is merely a preferred embodiment of this application and is not intended to limit this application in any other way. Any modifications or equivalent changes made based on the technical essence of this application shall still fall within the scope of protection claimed by this application.

Claims

1. A two-stage network-driven unsupervised adaptive segmentation method for cardiac structures, characterized in that, The method includes the following steps: S1. Collect labeled medical image data from the source domain and unlabeled medical image data from the target domain, and perform data preprocessing; S2. Construct and train an improved ISTN image translation network to convert source domain images into pseudo target domain images; S3. Construct an improved segmentation network for nnUNetv2, while integrating the Mean Teacher dual-branch architecture and dynamic pseudo-label strategy; S4. Combine the pseudo-target domain image with the source domain real label and the target domain unlabeled image as the training data of the segmentation network, perform iterative training, and check the convergence of the model through the validation set. S5. The final output is the trained segmentation network model.

2. The two-stage network-driven unsupervised adaptive segmentation method for cardiac structures according to claim 1, characterized in that, The improved ISTN image translation network in step S2 includes two generators, three discriminators, and two auxiliary segmentation networks. Both generators adopt an encoder-bottleneck layer-decoder structure, with only the bottleneck layer sharing weights. A content-guided attention fusion module is introduced between the encoder and decoder to alleviate the problems of blurred edges and contour distortion in the generated images. A feature enhancement upsampling module is added to the decoder to suppress artifacts during image reconstruction. An EfficientViMBlock module is embedded in the bottleneck layer to enhance the logical consistency of the global anatomical structure.

3. The two-stage network-driven unsupervised adaptive segmentation method for cardiac structures according to claim 1, characterized in that, The improved segmentation network in step S3 is based on nnUNetv2, which integrates the Mean Teacher dual-branch architecture and dynamic pseudo-label strategy. Specifically, the improvement of the segmentation network is achieved by introducing boundary feature alignment loss and uncertainty-guided region mixing module.

4. The two-stage network-driven unsupervised adaptive segmentation method for cardiac structures according to claim 2, characterized in that, The attention fusion module receives encoder features. With decoder features The specific process for adaptive feature fusion, using the input as data, is as follows: First, the encoder features and decoder features are added element-wise to obtain the initial fused features. ,right Perform global average pooling and combine it with a bottleneck structure of two 1x1 convolutions to generate channel attention weights. At the same time, Perform average pooling and max pooling along the channel dimension separately, concatenate the pooling results, and then generate spatial attention weights through a 7x7 convolution. ;Will and By performing broadcast summation, a coarse-grained spatial importance map is obtained. ; Then the initial fusion features Coarse-grained spatial importance map The data is concatenated along the channel dimension, and then channel shuffling, grouped convolution, and activation processing are performed sequentially to generate a fine-grained spatial importance map consistent with the input feature dimension. ; Finally, a detailed spatial importance map was used. As a weighting factor, the encoder features With decoder features Perform dynamic modulation and output the fused feature map. The calculation formula is: ； Here, ☉ represents element-wise multiplication.

5. The two-stage network-driven unsupervised domain adaptive segmentation method for cardiac structures according to claim 2, characterized in that, The specific operation flow of the feature enhancement upsampling module is as follows: First, the low-resolution feature map is received and divided into 8 sub-feature groups according to the channel dimension. A dual-branch parallel architecture is used to capture multi-scale information. The first branch uses 1×1 convolution combined with one-dimensional global average pooling in the horizontal and vertical directions to encode global information. The second branch uses 3×3 convolution to capture local spatial context. Then, through a cross-space learning mechanism, matrix multiplication and Softmax normalization are performed on the dual-branch output to generate spatial attention weights. The attention weights are multiplied element-wise with the feature groups and then upsampled by 2 times. At the same time, deep convolution is applied independently to each feature group to suppress noise and artifacts. After the feature groups are concatenated, cross-group information interaction is achieved through channel shuffling. Finally, feature fusion is completed by adjusting the number of channels using a 1×1 convolution.

6. The two-stage network-driven unsupervised domain adaptive segmentation method for cardiac structures according to claim 2, characterized in that, The EfficientViMBlock module comprises four components: a first deep convolutional layer, an HSM-SSD mixer, a second deep convolutional layer, and a feedforward network (FFN). Each component employs a LayerScale mechanism for feature fusion, with the fusion formula as follows: ； in, For the output of the current component, This is the output of the previous component. For learnable scaling parameters, For each component.

7. The two-stage network-driven unsupervised domain adaptive segmentation method for cardiac structures according to claim 3, characterized in that, The calculation process for the boundary feature alignment loss is as follows: First, Laplacian convolution is used to extract the boundary of the true label in the source domain. Then, the boundary is extracted after the Teacher branch prediction result is denoised by Gaussian-Laplacian operator. Then, the feature maps of each stage of the Student branch encoder are mapped to boundary feature maps, and the boundary dimensions are matched by interpolation; Finally, the boundary cross-entropy loss and the boundary Dice loss are summed to calculate the segmentation loss term for labeled data and the consistency constraint term for unlabeled data, respectively.

8. The two-stage network-driven unsupervised domain adaptive segmentation method for cardiac structures according to claim 3, characterized in that, The uncertainty-guided region hybrid module refers to calculating the entropy value based on the student and teacher model outputs to obtain the uncertainty map. : ； ； in, It is the output of the model. For class probabilities; The image is then divided into 64 regions. The three most uncertain regions from the student model's perspective are replaced with the three most certain regions from the teacher model's perspective to generate new samples. The quality of the pseudo-labels is optimized, and the generated new samples are input into the Student branch for training to enhance the learning of high-confidence features.

Images