Domain generalization method for medical image
By generating images with different domain styles during the training phase and calibrating them during the testing phase, the problem of domain offset in medical image segmentation is solved, achieving higher accuracy and robustness.
Patent Information
- Application Number
- CN202511175181.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-21
- Publication Date
- 2025-12-09
AI Technical Summary
The problem that existing medical image segmentation methods perform well during training but poorly during testing is mainly due to neglecting domain differences during the testing phase and failing to effectively utilize cross-domain distribution features.
During the training phase, a domain style enhancement module is introduced to generate images with different domain styles through Fourier transform. During the testing phase, a domain calibration module is introduced to calibrate the test sample images using multi-source amplitude prototype centroids, thereby reducing the difference between the target domain and the source domain.
It significantly improves the model's generalization ability in unknown domains, enhances the accuracy and robustness of medical image segmentation, and strengthens the model's sensitivity and adaptability to domain-invariant information.
Smart Images
Figure CN121095271A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical image segmentation technology, and in particular to a method for domain generalization of medical images. Background Technology
[0002] Medical images are produced by medical imaging equipment such as X-rays, computed tomography (CT), magnetic resonance imaging (MRI), and ultrasound scans. The quality and interpretation of medical images are crucial for disease diagnosis. Region of interest (ROI) segmentation in medical images is the most important foundation in medical image analysis. Accurate, robust, and rapid image segmentation is the most important step before subsequent steps such as quantitative analysis and 3D visualization, and it also lays the fundamental foundation for important clinical applications such as image-guided surgery, radiotherapy planning, and treatment evaluation.
[0003] Domain generalization (DG) is widely used in medical image segmentation to address the domain bias problem. Domain bias refers to a significant difference between the source and target domains in a machine learning task. This significant difference may cause the model to perform well during training but poorly in application. Most existing DG methods suffer from two significant drawbacks: First, these methods can only utilize information from the existing source domain and cannot extract discriminative information beyond that range. Second, existing methods primarily focus on improving generalization performance during training, neglecting the possibility of further narrowing the domain bias during testing. Summary of the Invention
[0004] The purpose of this invention is to provide a domain generalization method for medical images. During the training phase, images with different domain styles are obtained through a domain style enhancement module, thereby effectively utilizing the domain differentiation information in the original domain and the newly created domain to mine its potential cross-domain distribution features. During the testing phase, a domain calibration module is introduced. The domain calibration module uses multi-source amplitude prototype centroids to calibrate the test sample images, reducing the difference between the target domain and the source domain.
[0005] The above-mentioned technical objective of the present invention is achieved through the following technical solution: a domain generalization method for medical images, wherein a domain style enhancement module is introduced in the training phase, the domain style enhancement module performs Fourier transform and inverse transform on the sample images to obtain images with different domain styles; and a domain calibration module is introduced in the testing phase, the domain calibration module uses multi-source amplitude prototype centroids to calibrate the test sample images.
[0006] A further provision of the present invention is that the set containing k source domains is represented as: , It is the i-th image in the k-th source domain. The corresponding label; the target domain of an unlabeled image is represented as... The domain style enhancement module processes the sample images as follows: randomly selects sample images from the source domain Ds. Perform a Fourier transform on the sample image to obtain its amplitude and phase components. and These correspond to the amplitude and phase information of the image, respectively; another image is randomly selected from another source domain. The amplitude components are obtained by performing the same Fourier transform. Then use the following formula to extract the image. and images The amplitude information is merged: ,in, This represents the new amplitude component of the mixture. From A random variable for dynamic sampling is used to control the mixing ratio between the two components; a binary mask is introduced. Through inverse Fourier transform The combined amplitude components and the original phase component The process of converting the enhanced image to a Fourier transform is as follows: , where the image Include The semantic information, its structural information is and A mixture of.
[0007] A further provision of the present invention is that the domain encoding of a specific image can be identified as a k-dimensional vector, which satisfies... For Fourier transform-based image enhancement, its domain coding is defined as: ,in It is a hyperparameter that controls the ratio of original image domain coding to enhanced image domain coding. It is the corresponding random variable used in the amplitude mixing process.
[0008] A further provision of the present invention is that, in the domain style enhancement module, multi-scale encoder features are employed. As input, the features extracted from each layer of the encoder are fed into a global average pooling layer and a fully connected layer for aggregation. After processing by a softmax layer, the final domain prediction vector is obtained, represented as follows: , in, , , These represent the global average pooling layer, the fully connected layer, and the cascading operation, respectively. This represents the parameters of the fully connected layer; a single traditional convolutional layer is used as the domain-aware controller, mapping the domain encoding to variable parameters in the domain adaptation head. , When representing a dynamic convolutional layer, this layer enables the dynamic convolution to respond to specific domain information. The feature map generated by the decoder passes sequentially through the dynamic convolutional layer and the traditional convolutional layer to obtain the final segmentation prediction result, represented as... ,in, This represents a traditional convolutional layer. This indicates the encoder's output.
[0009] Among them, the domain adaptive head is the domain adaptive head.
[0010] A further provision of this invention is that, during training, the average value of each amplitude feature in the batch dimension is calculated to obtain its prototype centroid. The prototype centroid is then acquired and stored, and its corresponding frequency space prototype centroid can be represented as: ,in, It was estimated using all training samples in a single batch. For a specific batch i, to address the issue that a specific batch may not accurately reflect the prototype centroid of the global frequency space, the EMA algorithm is used to update the calibrated feature map: ,in This is the momentum parameter, used to establish the relationship between the amplitude of the calibrated feature map and the source domain image during inference during calibration. The input is an image that has already been calibrated during training. ,in It is a hyperparameter that controls the calibration intensity. This represents the amplitude of the source domain image.
[0011] A further provision of the present invention is that the domain calibration module uses a multi-source amplitude prototype centroid to calibrate the test sample image. Specifically, during the test, the amplitude prototype centroid is used to calibrate the test sample. To preserve the original image information, the final prediction result is obtained by integrating the original image and the calibrated image. ,in These are parameters for controlling amplitude calibration.
[0012] A further provision of the present invention is that, in the domain prediction task, the cross-entropy loss function is applied. ,in It's a domain tag. It's a prediction; for segmentation tasks, the cross-entropy loss function is applied. ,in, and Representing the segmentation prediction result and the ground truth, respectively, predictions are performed on both the original image and the enhanced image. This process yields... , , , Four indicators.
[0013] A further configuration of the present invention includes a teacher model and a student model. During training, the teacher model, updated with momentum, is used for consistency alignment, and parameters from the student model are received via exponential moving average. ,in , These are the parameters of the two models. The student model is optimized using stochastic gradient descent, while the teacher model does not pass any gradient information during backpropagation.
[0014] A further feature of this invention is that there is model consistency between the original image and the enhanced image output. A dual-path consistency constraint strategy is adopted to make the student model robust to domain shifts. The student output of the augmented image is consistent with the teacher output of the original image, and the student output of the original image is also consistent with the teacher output of the augmented image. KL divergence is used as the consistency metric. , ; All loss functions are combined into a final loss function. ,in , and These are hyperparameters used to balance the weights of the basic segmentation loss, domain classification loss, and consistency loss.
[0015] The beneficial effects of this invention are as follows: This invention proposes an advanced generalization method for medical images. This method generates diverse data with different domain attributes through Fourier transform-based data augmentation techniques, thereby effectively utilizing domain-discriminating information. Both the original image and the augmented image are concretized and applied, significantly improving the model's generalization ability to unknown domains. During the testing phase, this invention employs multi-source amplitude centroid alignment image styles to reduce domain differences. Attached Figure Description
[0016] Figure 1 This is a schematic diagram of the model architecture of the present invention. (a) Fourier-based Domain Style Enhancement Module (FADP). It adopts a bidirectional structure design, utilizing both raw and enhanced data for domain prediction. (b) Fourier-guided Domain Calibration Module (FGDC), which reduces domain discrepancies during inference.
[0017] Figure 2 It is based on Fourier transform to obtain images of different domain styles.
[0018] Figure 3 This is a visualization of the prediction results of the model of this invention and the results of six competing methods, as well as the actual values. The best results are displayed in color.
[0019] Figure 4 The t-SNE visualization results of unprocessed (left) and FADP-processed (right) features are shown, represented by dots of different colors.
[0020] Figure 5 This is an ablation study of the dual-path consistency strategy of this invention. The blue and orange bars represent the average DSC (%) of the model of this invention and the model without consistency loss, respectively. Detailed Implementation
[0021] A domain generalization method for medical images introduces a domain style enhancement module during the training phase. This module performs Fourier transform and its inverse transform on the sample images to obtain images with different domain styles. The specific process is as follows: A set containing k source domains is represented as: , It is the i-th image in the k-th source domain. The corresponding label; the target domain of an unlabeled image is represented as... The goal of this invention is to train a segmentation model on Ds so that it performs well on the target domain Dt.
[0022] The framework of this invention includes a teacher model and a student model. During training, the teacher model with momentum updates is used for consistency alignment, and parameters from the student model are received through exponential moving average. ,in , These are the parameters of the two models. The student model is optimized using stochastic gradient descent, while the teacher model does not pass any gradient information during backpropagation.
[0023] The domain style enhancement module processes the sample images as follows: randomly selects sample images from the source domain Ds. Perform a Fourier transform on the sample image to obtain its amplitude and phase components. and These correspond to the amplitude and phase information of the image, respectively; another image is randomly selected from another source domain. The amplitude components are obtained by performing the same Fourier transform. Then use the following formula to extract the image. and images The amplitude information is merged: ,in, This represents the new amplitude component of the mixture. From The random variable in the dynamic sampling is used to control the mixing ratio between the two components.
[0024] Introducing binary masks The cutoff frequency, or cutoff frequency, controls the amplitude of the low-frequency components that need to be fused in the amplitude spectrum. This is achieved through inverse Fourier transform. The combined amplitude components and the original phase component The process of converting the enhanced image to a Fourier transform is as follows: , where the image Include The semantic information, its structural information is and The mixture of Fourier transform and its inverse can both be efficiently performed using the Fast Fourier Transform algorithm.
[0025] Domain Discriminative Information Prediction: Due to the differences between the source and target domains, static encoder-decoder models trained solely on source domain data may not meet the requirements. To address this, this invention equips the model with a domain adaptation module, establishing connections between the model and data from different domains through dynamic convolutional structures and variable parameters. During training, the domain style enhancement module calculates the intrinsic domain information of the input data and feeds it into the domain-aware controller, thereby dynamically generating the parameters required for the domain adaptation head.
[0026] The domain encoding of a specific image can be identified as a k-dimensional vector that satisfies During training, the input image without Fourier transform is sampled from k source domains. Therefore, the true domain encoding clearly exhibits a one-hot encoding form. In existing technologies, the domain information carried by images that do not belong to any original source domain is ignored. In fact, the domain discrimination information of these fused images can be concretized and applied, thereby significantly improving the model's sensitivity to domain-invariant information. For Fourier transform-based enhanced images, its domain encoding is defined as: ,in It is a hyperparameter that controls the ratio of original image domain coding to enhanced image domain coding. This is because the present invention notes that the phase features of the image remain unchanged during the enhancement process, and these features contain structural information. Therefore, it is necessary to preserve the influence of high-level semantics when determining the domain coding. It is the corresponding random variable used in the amplitude mixing process.
[0027] In the domain style enhancement module, multi-scale encoder features are employed. As input, the features extracted from each layer of the encoder are fed into a global average pooling layer and a fully connected layer for aggregation. After processing by a softmax layer, the final domain prediction vector is obtained, represented as follows: ,in, , , These represent the global average pooling layer, the fully connected layer, and the cascading operation, respectively. This represents the parameters of the fully connected layer; a single traditional convolutional layer is used as the domain-aware controller, mapping the domain encoding to variable parameters in the domain adaptation head. , When representing a dynamic convolutional layer, this layer allows the dynamic convolution to respond to information from a specific domain. Finally, the feature map generated by the decoder is passed sequentially through the dynamic convolutional layer and the traditional convolutional layer to obtain the final segmentation prediction result, represented as... ,in, This represents a traditional convolutional layer. This represents the encoder's output. The final result is obtained by subtracting the domain discriminant components processed by dynamic convolution from the feature map after standard convolution.
[0028] In summary, this invention utilizes the domain discrimination information embedded in the encoder feature map to generate domain codes for each input image. The final... The value can be viewed as retaining the domain-invariant information while removing the domain-discriminating information.
[0029] During the testing phase, a domain calibration module is introduced, which uses the multi-source amplitude prototype centroid to calibrate the test sample image.
[0030] To reduce the discrepancy between test data and source data during inference, this invention employs a calibration feature map based on Fourier transform. First, during the training phase, the prototype centroid of the frequency space is determined from multiple source domains by calculating the average value of each amplitude feature in the batch dimension. Specifically, the amplitude components of each batch of images are acquired and stored; the corresponding prototype centroid of the frequency space can be represented as: .in, It was estimated using all training samples in a single batch. It should be noted that... This method is applicable to a specific batch i, and therefore may not accurately reflect the global frequency space prototype centroid. To address the issue that a specific batch may not accurately reflect the global frequency space prototype centroid, the EMA algorithm is used to update the calibrated feature map. ,in This is the momentum parameter, which is set to 0.99 in this invention. This makes the calibration feature map more robust in reflecting global information.
[0031] To establish the relationship between the amplitude of the calibration feature map and the source domain image during inference during calibration, the input image is an image that has already been calibrated during training: ,in It is a hyperparameter that controls the calibration intensity. This represents the amplitude of the source domain image.
[0032] The domain calibration module uses multi-source amplitude prototype centroids to calibrate the test sample images. Specifically, during the test, the amplitude prototype centroids are used to calibrate the test samples. To preserve the original image information, the final prediction result is obtained by integrating the original image and the calibrated image. ,in This is a parameter controlling the amplitude calibration. This invention sets it to 0.5 because it aims to avoid both complete loss of target amplitude information during testing and unnecessary distortion or artifacts. During training, this invention dynamically maintains a calibration amplitude that remains close to the centroid of the source domain style, enabling it to mitigate significant differences between the target and source domain data.
[0033] During the testing phase, this invention uses the EMA update mode, which enables the final centroid amplitude to robustly reflect the average style of the visible domain, while avoiding a large amount of computation.
[0034] Loss function: In domain prediction tasks, the cross-entropy loss function is applied. ,in It's a domain tag. It's a prediction; for segmentation tasks, the cross-entropy loss function is applied. ,in, and Representing the segmentation prediction result and the ground truth, respectively, predictions are performed on both the original image and the enhanced image. This process yields... , , , Four indicators.
[0035] There is model consistency between the original image and the enhanced image output. This invention employs a dual-path consistency constraint strategy to make the student model robust to domain shifts. Specifically, the student output of the augmented image is consistent with the teacher output of the original image, and vice versa. This design enables the model to learn domain-independent information through dual-path balanced learning. In particular, this invention uses KL divergence as a consistency metric. , Combine all loss functions into a final loss function. ,in , and These are hyperparameters used to balance the weights of the basic segmentation loss, domain classification loss, and consistency loss. In this invention, they are experimentally set to 1, 0.5, and 0.5.
[0036] Experimental Section This invention validates the proposed method in two medical segmentation tasks: joint segmentation of the optic cup and optic disc, and prostate segmentation. For each segmentation task, leave-one-domain validation is used to evaluate the performance of the domain-specific generalization methods—that is, training is performed on K-1 source domains, and evaluation is conducted on an unseen left-side target domain. Each domain is sequentially selected as the target domain for testing.
[0037] (1) Dataset and metrics OD / OC Segmentation Task: We obtained retinal fundus images from three publicly available datasets from four medical institutions: REFUGE, DrishtiGS, and RIM-ONE-r3. A simple UNet model was trained using a standard procedure to detect and crop 800×800 regions of interest (ROIs) around the optic disc to preprocess the fundus images. These cropped ROIs were then resized to 500×500 as network input. The larger ROI size relaxes positional bias constraints, ensuring coverage of all optic disc and cup regions even with data distribution shifts. The model performance was evaluated by calculating the average performance of the optic disc and cup region segmentation task. Each region was segmented into training and test sets according to the criteria used in previous studies. The dataset for the prostate segmentation task contained 116 T2-weighted MRI cases from six regions, collected from three publicly available datasets: NCI-ISBI13, I2CVB, and PROMISE12. All images were uniformly resized to 384×384 in the axial plane, and each data volume was standardized by subtracting the mean and dividing by the standard deviation.
[0038] The specific details of the datasets for the two tasks are shown in Table 1. Table 1. Dataset Descriptions for the Two Tasks
[0039] Segmentation performance was measured using the Dice similarity coefficient (DSC) and average surface distance (ASD), which characterize the accuracy of the predicted mask and boundary, respectively.
[0040] (2) Experiment content Our model was trained using the PyTorch framework on a workstation equipped with an NVIDIA RTX 3090 graphics card. We chose a batch size of 16 to balance network complexity with image size. The training set was augmented using basic data augmentation techniques such as random flipping, rotation, scaling, and cropping. Stochastic gradient descent with a momentum of 0.99 was used as the optimizer, with an initial learning rate of [missing information]. Set the learning rate to 0.001. The learning rate follows a polynomial decay rule. Adjustments are made, where t represents the current training epoch, and T is the maximum epoch (set to 300). The momentum parameter m of the teacher model is set to 0.9995.
[0041] (3) Experimental results and analysis Our comparison method includes baseline and state-of-the-art domain generalization models, as shown below: DCAC: A model containing two dynamically convolutional modules for capturing domain discrimination information embedded in the encoder feature map.
[0042] RAM: A data augmentation-based approach that utilizes low-frequency information and leverages image restoration to learn more robust feature representations.
[0043] DOCR: A data augmentation-based approach that includes a high-frequency reconstruction module. Note that we have removed the involvement of the target domain data in the original method.
[0044] FedDG: A federated learning approach that uses distribution information passed through the frequency space and leverages contextual learning from each local client.
[0045] DoFE: A domain-invariant feature learning method that dynamically enriches image features by introducing additional domain prior knowledge, thereby making semantic features more discriminative.
[0046] DDG: A data augmentation-based method that dynamically adjusts parameters and simulates the style of the target domain.
[0047] In addition, we compared the proposed model with a baseline setting where the model was trained on aggregated data from all source domains and tested directly on an unseen target domain.
[0048] Table 2 presents the OD / OC segmentation results. We can observe that, compared to baseline methods, domain transfer techniques have shown significant improvements in all DG methods designed to address this issue. Our method shows the most outstanding performance, achieving an average improvement of 3.48% in the dice coefficient metric and an average improvement of 2.56 voxels in the ASD metric. Unlike DCAC, we not only utilize known domain information but also dynamically mine hidden composite domain information, thereby enhancing the model's ability to perceive sample domain attributes. Furthermore, compared to the RAM method, which also employs Fourier transform enhancement, our method achieves a significant improvement by effectively utilizing image domain attributes. It is noteworthy that the performance improvement is more significant when the model's data source is richer. This may be because the model relies more heavily on data to enhance its adaptability to domain attributes.
[0049] The prostate MRI segmentation results are shown in Table 3. The experimental results show that the baseline method without domain generalization performs poorly due to the distributional differences between the source and target data. Our method achieved the highest Dice coefficient and ASD score among similar methods. Its average Dice coefficient reached 89.64%, and the ASD value was 1.20 voxels, both superior to other domain generalization methods. Specifically, compared to the baseline method, our method improved the overall Dice coefficient by 5.11% and reduced the ASD value by 1.41 voxels, demonstrating a significant advantage.
[0050] In addition, we Figure 3 The final segmentation results are visualized, showing predictions that most closely approximate the ground truth annotations. Our model generates segmentation results with the highest similarity to ground truth annotations, exhibiting clearer boundaries and shapes. This demonstrates the effectiveness of our method compared to previous state-of-the-art methods in general medical image segmentation tasks across two different imaging modalities.
[0051] Table 2 shows the performance of different methods on the OD / OC segmentation task, with the best results highlighted in bold.
[0052] Table 3. Performance of different methods on prostate MRI segmentation tasks; the best results are highlighted in bold.
[0053] (4) Generalization analysis of FADP module To more intuitively demonstrate the performance of the FADP module, we Figure 4 The image visualization results of the fundus dataset are presented in the image t-SNE. The left side shows the original distribution information of different domains in the fundus dataset, and the significant separation of image features in each domain can be clearly seen. This feature differentiation may lead to overfitting risk when the model is trained in the source domain, thus affecting the generalization ability of the target domain. However, by applying the FADP module, we can effectively reduce the domain differences and demonstrate a representation ability with domain invariance.
[0054] (5) Ablation research To verify the effectiveness of each component of the model, we compared it with a variant using only a single module. The experimental results are shown in Tables 4 and 5. These experimental data demonstrate that each part of the model contributes to the overall performance improvement. It is worth noting that our FGDC module performs poorly in certain domains, which may be related to the sensitivity of the images themselves.
[0055] In these specific domains, artifacts generated during amplitude mixing are particularly prominent. However, this effect is mitigated by Fourier-based enhancement techniques after coupling, thereby improving system performance. During training, we employ dual-path consistency loss to enhance the model's resistance to domain shifts. Figure 5 As shown, we evaluated the impact of the consistency loss on the fundus OD / OC segmentation task. The results show an improvement in average performance, highlighting the effectiveness of the consistency loss in improving the model's generalization ability and enhancing its robustness to domain shifts.
[0056] Experimental results on two public datasets demonstrate that the method of this invention exhibits strong generalization ability and industry-leading performance.
[0057] Table 4 Performance of the baseline model, the model of this invention, and two variants in OD / OC segmentation.
[0058] Table 5 Performance of the baseline model, the model of this invention, and two variants in prostate segmentation.
Claims
1. A domain generalization method for medical images, characterized in that: During the training phase, a domain style enhancement module is introduced. The domain style enhancement module performs Fourier transform and inverse transform on the sample images to obtain images with different domain styles. During the testing phase, a domain calibration module is introduced, which uses the multi-source amplitude prototype centroid to calibrate the test sample image.
2. The domain generalization method for medical images according to claim 1, characterized in that: A set containing k source domains is represented as: , It is the i-th image in the k-th source domain. The corresponding label; the target domain of an unlabeled image is represented as... ; The domain style enhancement module processes the sample image as follows: from the source domain D... s Randomly selected sample images Perform a Fourier transform on the sample image to obtain its amplitude and phase components. and These correspond to the amplitude and phase information of the image, respectively; Randomly select another image from another source domain. The amplitude components are obtained by performing the same Fourier transform. Then use the following formula to extract the image. and images The amplitude information is merged: ,in, This represents the new amplitude component of the mixture. From A random variable for dynamic sampling is used to control the mixing ratio between two components; a binary mask is introduced. Through inverse Fourier transform The combined amplitude components and the original phase component The process of converting the enhanced image to a Fourier transform is as follows: , where the image Include The semantic information, its structural information is and A mixture of.
3. The domain generalization method for medical images according to claim 2, characterized in that: The domain encoding of a specific image can be identified as a k-dimensional vector that satisfies ; For Fourier transform-based image enhancement, its domain coding is defined as: ,in It is a hyperparameter that controls the ratio of original image domain coding to enhanced image domain coding. It is the corresponding random variable used in the amplitude mixing process.
4. The domain generalization method for medical images according to claim 3, characterized in that: In the domain style enhancement module, multi-scale encoder features are employed. As input, the features extracted from each layer of the encoder are fed into a global average pooling layer and a fully connected layer for aggregation. After processing by a softmax layer, the final domain prediction vector is obtained, represented as follows: ,in, , , These represent the global average pooling layer, the fully connected layer, and the cascading operation, respectively. This represents the parameters of the fully connected layer; a single traditional convolutional layer is used as the domain-aware controller, mapping the domain encoding to variable parameters in the domain adaptation head. , When representing a dynamic convolutional layer, this layer enables the dynamic convolution to respond to specific domain information. The feature map generated by the decoder passes sequentially through the dynamic convolutional layer and the traditional convolutional layer to obtain the final segmentation prediction result, represented as... ,in, This represents a traditional convolutional layer. This indicates the encoder's output.
5. The domain generalization method for medical images according to claim 4, characterized in that: During training, the average value of each amplitude feature in the batch dimension is calculated to obtain its prototype centroid. The amplitude components of each batch of images are acquired and stored, and the corresponding frequency space prototype centroid can be represented as: ,in, It was estimated using all training samples in a single batch. For a specific batch i, to address the issue that a specific batch may not accurately reflect the prototype centroid of the global frequency space, the EMA algorithm is used to update the calibrated feature map: ,in This is the momentum parameter, used to establish the relationship between the amplitude of the calibrated feature map and the source domain image during inference during calibration. The input is an image that has already been calibrated during training. ,in It is a hyperparameter that controls the calibration intensity. This represents the amplitude of the source domain image.
6. The domain generalization method for medical images according to claim 5, characterized in that: The domain calibration module uses multi-source amplitude prototype centroids to calibrate the test sample images. Specifically, during the test, the amplitude prototype centroids are used to calibrate the test samples. To preserve the original image information, the final prediction result is obtained by integrating the original image and the calibrated image. ,in These are parameters for controlling amplitude calibration.
7. The domain generalization method for medical images according to claim 6, characterized in that: In domain prediction tasks, the cross-entropy loss function is applied. ,in It's a domain tag. It's a prediction; for segmentation tasks, the cross-entropy loss function is applied. ,in, and Representing the segmentation prediction result and the ground truth, respectively, predictions are performed on both the original image and the enhanced image. This process yields... , , , Four indicators.
8. The domain generalization method for medical images according to claim 7, characterized in that: It includes a teacher model and a student model. During training, the teacher model, which uses momentum updates, is used for consistency alignment, while parameters from the student model are received through an exponential moving average. ,in , These are the parameters of the two models. The student model is optimized using stochastic gradient descent, while the teacher model does not pass any gradient information during backpropagation.
9. The domain generalization method for medical images according to claim 8, characterized in that: There is model consistency between the original image and the enhanced image output. A dual-path consistency constraint strategy is adopted to make the student model robust to domain shifts. The student output of the augmented image is consistent with the teacher output of the original image, and the student output of the original image is also consistent with the teacher output of the augmented image. KL divergence is used as the consistency metric. , Combine all loss functions into a final loss function. ,in , and These are hyperparameters used to balance the weights of the basic segmentation loss, domain classification loss, and consistency loss.