Cross-domain few-sample brain tumor MRI segmentation method without participation of source domain data

Abstract

The invention discloses a cross-domain few-sample brain tumor MRI (Magnetic Resonance Imaging) segmentation method without participation of source domain data, aiming at a scene in which source domain original data is prohibited to be shared clinically and target domain labels are few, the method only utilizes a public pre-training backbone and 1-5 hospital label MRI: firstly, dual random enhancement is performed in pixel and feature space to expand domain coverage; simultaneously correcting support and query features by using a single saliency heat map to realize tumor region alignment; then, through combined constraint of foreground consistency comparison loss and triple semantic discrimination loss, intra-class compactness and inter-class separability are improved; and during reasoning, performing multi-level similarity fusion on a query image, binarizing by using an Otsu threshold value and a maximum value of a global mean value, and outputting a final segmentation result. According to the method, the privacy compliance requirement that the data cannot be out of the domain is met, rapid deployment can be carried out in a new hospital, the labeling cost is remarkably reduced, and high precision is kept.

Description

Technical Field

[0001] This invention relates to a cross-domain few-sample MRI segmentation method for brain tumors that does not require source domain data, and belongs to the field of medical image segmentation technology. Background Technology

[0002] Automated segmentation of brain tumors on MRI relies on a large amount of labeled data, but domain differences across hospitals, equipment, and populations cause a sharp drop in model performance upon deployment. Existing cross-domain few-shot segmentation (CD-FSS) methods generally assume access to the original source domain images and mitigate domain shift through source-target joint training or style transfer. However, clinical data privacy regulations explicitly prohibit the export or sharing of patients' original images; newly established hospitals often only have publicly available pre-trained weights and a very small number of annotations (1–5 cases) from their own hospital, creating a new scenario of "no source data and few samples." In this scenario, traditional CD-FSS loses its data foundation, resulting in a significant performance degradation, and may even prevent training from starting. Therefore, how to achieve high-precision brain tumor MRI segmentation using only publicly available backbones and a very small number of hospital annotations without touching the original source domain data has become a pressing technological gap that needs to be addressed. Summary of the Invention

[0003] The present invention provides a cross-domain few-sample MRI segmentation method for brain tumors that does not require source domain data to solve the problems existing in the prior art.

[0004] The technical solutions adopted in this invention are as follows:

[0005] A cross-domain few-sample MRI segmentation method for brain tumors that does not require source domain data includes the following steps:

[0006] (a) Input the support images and query images in the image set into the same enhancement-feature extraction module. In the enhancement-feature extraction module, each image is first randomly enhanced in the pixel space, and then the initial features are extracted by the backbone network with shared weights. Subsequently, the initial features are randomly enhanced again in the feature space, and finally the support features and query features are output.

[0007] (b) The supporting features and query features are fed into the heatmap-guided structure alignment module to generate a spatial saliency heatmap. The supporting features are first weighted by the saliency of the heatmap, and then fused with the query features to output the structure-aligned supporting features and query features.

[0008] (c) The structure-aligned support features and query features are fed into the semantic consistency alignment module to calculate the foreground consistency contrast loss and triple semantic discrimination loss. The two types of losses are weighted and summed using preset weights. The total loss constrains the network parameters to complete iterative training and obtain the trained SHCNet model.

[0009] (d) The trained SHCNet model is used to perform multi-level similarity inference on the new query image to obtain the fusion prediction map. Then, the maximum value of the Otsu threshold and the global mean is selected as the segmentation threshold. The fusion prediction map is binarized and the final brain tumor binary segmentation result is output.

[0010] Furthermore, the random enhancement of the pixel space is at least one of the following operations: rotation, scaling, noise addition, blurring, and color perturbation; the backbone network with shared weights is a ResNet series network or a DinoV2 network.

[0011] Furthermore, the image set includes a support set and a query set; the support set consists of at least one labeled support image of a brain tumor mask, and the query set consists of unlabeled query images to be segmented;

[0012] The initial features are then augmented randomly in the feature space based on channel statistics, using the EnhanceStyle module. The specific process is as follows:

[0013] (1) From the original initial features of the support set and initial features after pixel enhancement Two feature samples were randomly sampled. , ,Will Replace with A copy is obtained to get the feature sample group. ;

[0014] (2) Based on Distributed sampling perturbation coefficient Calculate the feature sample groups respectively and Channel mean with standard deviation The mixture statistic is calculated using the following formula:

[0015] ;

[0016] (3) Based on the formula Complete the characteristic style perturbation, among which The final output supports both the features and the query features.

[0017] Furthermore, when generating the spatial saliency heatmap, semantic compression and channel transformation are first performed on the supporting features and query features, and then spatial saliency priors are constructed. The specific process is as follows:

[0018] The supporting features and query features are fed into a feature adaptation layer consisting of two cascaded 1×1 convolutions. After the first convolution, batch normalization and ReLU activation function are applied to complete semantic compression and channel transformation, resulting in adapted features.

[0019] The adapted feature map is divided into four regions, and the pixel with the largest response value in each region is selected as the local saliency center. Expanding in four directions until the neighborhood mean is below the threshold of 128, the radius of the significant region is obtained. ;

[0020] For each saliency center Generate a two-dimensional Gaussian kernel The final spatial saliency heatmap is obtained by taking the maximum value of all Gaussian kernel responses per pixel. .

[0021] Furthermore, in step (b), the specific method for weighting the supporting features based on the spatial saliency heatmap and fusing them with the query features is as follows:

[0022] According to the formula Generate a feature fusion map A, where To support features, For query features;

[0023] Structure alignment features are obtained by element-wise multiplication and feature superposition, i.e. ,in For element-wise multiplication, this is used as the supporting and query features of the output structure after alignment.

[0024] Furthermore, in step (c), the calculation of the foreground consistency contrast loss requires first constructing feature prototypes under different enhanced views, and then constraining semantic consistency through cosine similarity, specifically as follows:

[0025] Original support foreground features after structural alignment Enhance support for future features Compared with the original supporting background features Calculate the global average pooling separately to obtain the corresponding feature prototypes;

[0026] The loss value is calculated using the following formula to maximize foreground semantic consistency across different views:

[0027] .

[0028] Furthermore, in step (c), the triple semantic discriminative loss is constructed as follows:

[0029] To query foreground features Anchor point, supporting foreground features Positive examples, supporting background features For negative examples, the loss formula is:

[0030] ,

[0031] Among them, boundary hyperparameters .

[0032] Furthermore, in step (c),

[0033] The default weighting rule is: weight of foreground consistency contrast loss. Weights of triple semantic discriminative loss The total loss formula is: The network parameters are then updated using this total loss through backpropagation to complete iterative training.

[0034] Furthermore, in step (d), the specific process for obtaining the fusion prediction graph through multi-level similarity reasoning is as follows:

[0035] For each feature layer m, according to the formula Preliminary prediction within the computational layer, where d is the feature dimension. To support masking, the preliminary prediction results of all feature layers are upsampled to the original image size using bilinear interpolation, and then fused using the mean. ,in The feature layer number is used as the fusion prediction map.

[0036] Furthermore, in step (d), the specific logic of the binarization process is as follows: The fused prediction image is... The response value is greater than The region with the highest response value is identified as the brain tumor region, and the region with the highest response value is identified as the background region. Finally, a 0-1 binary brain tumor segmentation mask is output.

[0037] The present invention has the following beneficial effects:

[0038] The entire training and inference process does not require access to or transmission of the original source domain images, meeting the privacy compliance requirements of clinical data not leaving the domain or country. It can be deployed directly in newly built hospitals, and segmentation can be started with only 1-5 annotated MRI images from the hospital, significantly reducing the cost of manual delineation, and is suitable for rare cases or rapid deployment scenarios.

[0039] Random perturbations in both pixel and feature spaces effectively offset appearance differences across devices, sequences, and populations, reducing false negatives / positives caused by domain drift. The same saliency heatmap is used to simultaneously correct support and query features, enabling the model to focus on the tumor region and reducing boundary leakage caused by location shifts. Foreground consistency contrast loss and triplet discrimination loss are jointly optimized to maintain tumor feature stability under different enhancement perspectives while increasing the distance between the tumor and background, improving inter-class separability. During the inference phase, only forward similarity calculation and Otsu-Mean thresholding are performed, requiring no iteration or additional post-processing. A single GPU can complete single-case segmentation in seconds, suitable for routine clinical workflows. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the overall structure of the SHCNet model.

[0041] Figure 2 This is a detailed structural diagram of the heatmap-guided structure alignment module. Detailed Implementation

[0042] The invention will now be further described with reference to the accompanying drawings.

[0043] This invention presents a cross-domain, few-sample MRI segmentation method for brain tumors that does not require source domain data. Based on the SHCNet model, this method comprises three core functional modules: enhancement-feature extraction, heatmap-guided structure alignment, and semantic consistency alignment. It also incorporates a multi-level similarity inference process and an Otsu-means dual-threshold segmentation inference flow. The overall technical architecture is as follows: Figure 1 As shown, the model completes iterative optimization of network parameters during the training phase and achieves accurate segmentation of brain tumors in newly queried MRI images during the inference phase, without the need for source domain data, thus meeting the practical needs of clinical data privacy and cross-institutional data isolation.

[0044] First, prepare the image set required for the experiment. This image set includes a support set and a query set. The support set consists of at least one labeled support image of a brain tumor mask, and the query set consists of unlabeled query images to be segmented. The support images and query images in the image set are input into the same enhancement-feature extraction module for feature processing. Within the enhancement-feature extraction module, each image undergoes random enhancement operations in pixel space. These random enhancements include at least one of the following: rotation, scaling, noise addition, blurring, and color perturbation. After pixel-level enhancement, the image is fed into a shared-weight backbone network to extract initial features. This backbone network can be a ResNet series network or a DinoV2 network.

[0045] After obtaining the initial features, it is necessary to perform random enhancement on the initial features again in the feature space. This random enhancement of the feature space is implemented through the EnhanceStyle module, and the specific process is as follows:

[0046] First, starting with the original initial features of the support set. and initial features after pixel enhancement Two feature samples were randomly sampled. , ,Will Replace with A copy is obtained to get the feature sample group. ;

[0047] Then, based on the Beta distribution, sample the perturbation coefficients. (Set in this embodiment) ), calculate the feature sample groups respectively and Channel mean with standard deviation The mixture statistic is calculated using the following formula:

[0048] ;

[0049] in , These represent the mean and standard deviation of each channel in the spatial dimension, respectively. , Where c, h, and w represent the channel, height, and width, respectively. Statistics are performed on each channel along the spatial dimensions. , These are the mixed mean and standard deviation, used to construct new feature styles.

[0050] Finally, based on the formula Complete the characteristic style perturbation, among which The final output supports both the features and the query features.

[0051] Combination Figure 2 The obtained supporting features and query features are then fed into the heatmap-guided structure alignment module. First, a spatial saliency heatmap is generated. During generation, semantic compression and channel transformation are performed on the supporting and query features. Specifically, the supporting and query features are fed into a feature adaptation layer consisting of two cascaded 1×1 convolutions. After the first convolution, batch normalization and a ReLU activation function are applied to complete semantic compression and channel transformation, resulting in the adapted features. The transformation process of this feature adaptation layer is defined as follows:

[0052] ,

[0053] in , It is a 1×1 convolution kernel. The features will be processed through max pooling and average pooling, respectively, and concatenated with the heatmap for the next round of convolution operations. This is used to extract spatial priors to achieve consistent alignment, and finally obtain the final result. . Figure 2 middle, For input features, It is an intermediate feature.

[0054] After feature fitting is completed, spatial saliency prior is constructed. The fitted feature map is first divided into four regions on average, and the pixel with the largest response value in each region is selected as the local saliency center. Expanding in four directions until the neighborhood mean is below the threshold of 128, the radius of the significant region is obtained. The calculation method is as follows:

[0055] ,

[0056] Then for each saliency center Generate the corresponding two-dimensional Gaussian kernel :

[0057] ,

[0058] All Gaussian responses are merged based on the maximum pixel value to form the final heatmap. : .

[0059] After obtaining the spatial saliency heatmap, the supporting features are first weighted saliency using this heatmap, and then fused with the query features. The specific method is as follows: [Formula omitted]. Generate a feature fusion map A, where To support features, To query features; then, by element-wise multiplication and superposition with the features, we obtain structure-aligned features, i.e. ,in This represents element-wise multiplication, which outputs the structure-aligned supporting features and query features.

[0060] The structure-aligned support features and query features are fed into the semantic consistency alignment module. First, the foreground consistency contrast loss is calculated. During the calculation, feature prototypes under different augmented views need to be constructed first. Then, semantic consistency is constrained by cosine similarity. Specifically, this involves adjusting the original support and foreground features after structure alignment. Enhance support for future features Compared with the original supporting background features Calculate global average pooling separately to obtain the corresponding feature prototypes; then calculate the loss value according to the following formula to maximize the semantic consistency of the foreground under different views:

[0061] ,

[0062] This module also requires calculating the triple semantic discrimination loss, which is constructed using query foreground features. Anchor point, supporting foreground features Positive examples, supporting background features For negative examples, the loss formula is:

[0063] ,

[0064] Among them, boundary hyperparameters .

[0065] After calculating the two types of losses, a weighted sum of the two types of losses is performed using preset weights. In this embodiment, the preset weights are determined by the weight of the foreground consistency contrast loss. Weights of triple semantic discriminative loss The total loss formula is: The total loss is then used to backpropagate and update the network parameters to complete iterative training, resulting in the trained SHCNet model.

[0066] After the model training is completed, the trained SHCNet model is used to segment new query images. First, multi-level similarity inference is performed to obtain a fused prediction map. The specific process is as follows: for each feature layer m, according to the formula:

[0067] Preliminary prediction within the computational layer, where d is the feature dimension. To support masks.

[0068] Then, the preliminary prediction results of all feature layers are upsampled to the original image size using bilinear interpolation, and fused by mean to obtain the final image. ,in The feature layer number is used as the fusion prediction map.

[0069] After obtaining the fused prediction map, the maximum value of the Otsu threshold and the global mean is selected as the segmentation threshold. The fused prediction map is then binarized. The specific logic is as follows: the fused prediction map... The response value is greater than The region is identified as the brain tumor region, and the region with a response value lower than the threshold is identified as the background region. Finally, a 0-1 binary brain tumor segmentation mask is output to complete the segmentation of the brain tumor by MRI.

[0070] To verify the effectiveness of this method, comparative experiments were conducted on three datasets: BraTS2020, BraTS2021, and BraTSAfrica. The comparison methods included mainstream cross-domain few-shot segmentation methods that do not require source data, such as PATNet, RestNet, TGCM, ABCDFSS, and MedXNet. The core evaluation metrics were DSC (similarity coefficient), FB-IoU (foreground-background intersection-over-union ratio), and HD95 (95% Hausdorff distance). Specific experimental results are shown in Tables 1, 2, and 3 below.

[0071] Table 1. Quantization segmentation results of different CD-FSS methods on the BraTS2020 dataset.

[0072]

[0073] Table 2. Quantization segmentation results of different CD-FSS methods on the BraTS2021 dataset.

[0074]

[0075] Table 3. Quantization segmentation results of different CD-FSS methods on the BraTSAfrica dataset.

[0076]

[0077] The experimental data above show that, using ResNet-50 as the backbone network, the average DSC of this method on the BraTS2020 dataset under 1-shot and 5-shot conditions reached 76.22% and 80.36%, respectively, representing improvements of 12.54 and 16.23 percentage points compared to ABCDFSS. On the BraTS2021 dataset, the average DSC under 1-shot and 5-shot conditions were 77.91% and 79.70%, respectively, representing improvements of 14.87 and 16.65 percentage points compared to ABCDFSS. On the BraTSAfrica dataset, which exhibits the most severe style shift, the average DSC under 1-shot and 5-shot conditions were 74.74% and 76.75%, respectively, representing improvements of 8.27 and 10.93 percentage points compared to ABCDFSS. Furthermore, this method achieved the lowest HD95 across all datasets, reaching as low as 6.33 mm, demonstrating excellent boundary awareness and cross-domain robustness.

[0078] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements without departing from the principle of the present invention, and these improvements should also be considered within the scope of protection of the present invention.

Claims

1. A method for cross-domain few-sample MRI segmentation of brain tumors without the need for source domain data, characterized in that: Includes the following steps: (a) Input the support images and query images in the image set into the same enhancement-feature extraction module. In the enhancement-feature extraction module, each image is first randomly enhanced in the pixel space, and then the initial features are extracted by the backbone network with shared weights. Subsequently, the initial features are randomly enhanced again in the feature space, and finally the support features and query features are output. (b) The supporting features and query features are fed into the heatmap-guided structure alignment module to generate a spatial saliency heatmap. The supporting features are first weighted by the saliency of the heatmap, and then fused with the query features to output the structure-aligned supporting features and query features. (c) The structure-aligned support features and query features are fed into the semantic consistency alignment module to calculate the foreground consistency contrast loss and triple semantic discrimination loss. The two types of losses are weighted and summed using preset weights. The total loss constrains the network parameters to complete iterative training and obtain the trained SHCNet model. (d) The trained SHCNet model is used to perform multi-level similarity inference on the new query image to obtain the fusion prediction map. Then, the maximum value of the Otsu threshold and the global mean is selected as the segmentation threshold. The fusion prediction map is binarized and the final brain tumor binary segmentation result is output.

2. The cross-domain few-sample MRI segmentation method for brain tumors without the need for source domain data as described in claim 1, characterized in that: The random enhancement of the pixel space is at least one of the following operations: rotation, scaling, noise addition, blurring, and color perturbation; the backbone network with shared weights is a ResNet series network or a DinoV2 network.

3. The cross-domain few-sample MRI segmentation method for brain tumors without the need for source domain data as described in claim 1, characterized in that: The image set includes a support set and a query set; the support set consists of at least one labeled support image of a brain tumor mask, and the query set consists of unlabeled query images to be segmented; The initial features are then augmented randomly in the feature space based on channel statistics, using the EnhanceStyle module. The specific process is as follows: (1) From the original initial features of the support set and initial features after pixel enhancement Two feature samples were randomly sampled. , ,Will Replace with A copy is obtained to get the feature sample group. ; (2) Based on Distributed sampling perturbation coefficient Calculate the feature sample groups respectively and Channel mean with standard deviation The mixture statistic is calculated using the following formula: ； (3) Based on the formula Complete the characteristic style perturbation, among which The final output supports both the features and the query features.

4. The cross-domain few-sample MRI segmentation method for brain tumors without the need for source domain data as described in claim 1, characterized in that: When generating the spatial saliency heatmap, semantic compression and channel transformation are first performed on supporting features and query features, and then spatial saliency priors are constructed. The specific process is as follows: The supporting features and query features are fed into a feature adaptation layer consisting of two cascaded 1×1 convolutions. After the first convolution, batch normalization and ReLU activation function are applied to complete semantic compression and channel transformation, resulting in adapted features. The adapted feature map is divided into four regions, and the pixel with the largest response value in each region is selected as the local saliency center. Expanding in four directions until the neighborhood mean is below the threshold of 128, the radius of the significant region is obtained. ; For each saliency center Generate a two-dimensional Gaussian kernel The final spatial saliency heatmap is obtained by taking the maximum value of all Gaussian kernel responses per pixel. .

5. The cross-domain few-sample MRI segmentation method for brain tumors without the need for source domain data as described in claim 1, characterized in that: In step (b), the specific method for weighting the supporting features with spatial saliency heatmaps and fusing them with the query features is as follows: According to the formula Generate a feature fusion map A, where To support features, For query features; Structure alignment features are obtained by element-wise multiplication and feature superposition, i.e. ,in For element-wise multiplication, this is used as the supporting and query features of the output structure after alignment.

6. The cross-domain few-sample MRI segmentation method for brain tumors without the need for source domain data as described in claim 1, characterized in that: In step (c), the calculation of foreground consistency contrast loss requires first constructing feature prototypes under different enhanced views, and then constraining semantic consistency through cosine similarity, specifically: Original support foreground features after structural alignment Enhance support for future features Compared with the original supporting background features Calculate the global average pooling separately to obtain the corresponding feature prototypes; The loss value is calculated using the following formula to maximize foreground semantic consistency across different views: 。 7. The cross-domain few-sample MRI segmentation method for brain tumors without the need for source domain data as described in claim 1, characterized in that: In step (c), the triple semantic discriminant loss is constructed as follows: To query foreground features Anchor point, supporting foreground features Positive examples, supporting background features For negative examples, the loss formula is: ， Among them, boundary hyperparameters .

8. The cross-domain few-sample MRI segmentation method for brain tumors without the need for source domain data as described in claim 1, characterized in that: In step (c), The default weighting rule is: weight of foreground consistency contrast loss. Weights of the triple semantic discriminative loss The total loss formula is: The network parameters are then updated using this total loss through backpropagation to complete iterative training.

9. The cross-domain few-sample MRI segmentation method for brain tumors without the need for source domain data as described in claim 1, characterized in that: In step (d), the specific process for obtaining the fusion prediction graph through multi-level similarity reasoning is as follows: For each feature layer m, according to the formula Preliminary prediction within the computational layer, where d is the feature dimension. To support masking, the preliminary prediction results of all feature layers are upsampled to the original image size using bilinear interpolation, and then fused using the mean. ,in The feature layer number is used as the fusion prediction map.

10. The cross-domain few-sample MRI segmentation method for brain tumors without the need for source domain data as described in claim 1, characterized in that: In step (d), the specific logic of binarization is as follows: The fused prediction image is... The response value is greater than The region with the highest response value is identified as the brain tumor region, and the region with the highest response value is identified as the background region. Finally, a 0-1 binary brain tumor segmentation mask is output.

Images