An unsupervised cross-domain brain tumor segmentation method for low-quality MRI

Abstract

The application discloses a kind of low-quality MRI-oriented unsupervised cross-domain brain tumor segmentation methods.It includes CDGM-Net, the CDGM-Net includes three collaborative components: initial mask generation module IMG, frequency decomposition matching module FDM and frequency cross fusion module FCM;Specifically comprising the following steps: step S1.MRI image input and feature extraction;Step S2. Generate initial mask, obtain the initial foreground mask of query image feature;Step S3. Frequency decomposition, i.e. the first stage of FDM module;Step S4. Frequency distribution matching, i.e. the second stage of FDM module, obtain the low-frequency feature of aligned query image and aligned query high-frequency image feature;Step S5. Frequency cross fusion, i.e. FCM module: generate final brain tumor segmentation mask.The advantage is: in low-quality MRI scene, noise interference can be more effectively suppressed and the structure consistency of tumor region is maintained, the effectiveness and practical value of frequency domain alignment and cross-frequency fusion strategy in cross-dataset brain tumor segmentation task are verified.

Description

Technical Field

[0001] This invention relates to a deep learning-based cross-domain brain tumor segmentation method, and more particularly to an unsupervised cross-domain brain tumor segmentation method for low-quality MRI. Background Technology

[0002] Brain tumor segmentation plays a crucial role in clinical diagnosis, treatment planning, and prognostic assessment. With the rapid development of deep learning technology, significant progress has been made on well-constructed benchmark datasets. However, applying these models to real-world clinical settings remains challenging. In particular, MRI data acquired from different hospitals, scanners, and imaging protocols often exhibit significant quality variations, including reduced spatial resolution, increased noise levels, and decreased tissue contrast. These variations are particularly pronounced in resource-scarce regions due to poor imaging condition control and inherent limitations in acquisition quality. This quality degradation severely impairs model performance when models trained on high-quality MRI datasets (e.g., BraTS 2021 / 2020) are directly applied to low-quality MRI data (e.g., BraTS Africa), significantly limiting their clinical application. Low-quality MRI images typically feature strong noise interference, signal attenuation, and blurred tumor boundaries, hindering accurate differentiation between pathological and background tissues and significantly reducing cross-dataset generalization ability.

[0003] To address the distributional disparities between datasets, a significant amount of research has focused on cross-domain medical image segmentation, with domain adaptation (DA) and domain generalization (DG) becoming mainstream paradigms. DA methods typically rely on access to target domain samples during training and attempt to reduce inter-domain discrepancies through adversarial learning or statistical alignment. However, in real-world clinical scenarios—especially in resource-scarce regions—reliable target domain annotations are often difficult to obtain or prohibitively expensive, hindering the practical application of such methods. In contrast, DG and cross-domain few-shot segmentation methods aim to improve robustness to unknown domains without retraining the target domain, but most existing methods operate primarily in spatial or pixel feature spaces. Consequently, these methods remain highly sensitive to variations in image quality, noise patterns, and resolution, leading to unstable foreground prototypes and decreased boundary prediction accuracy, particularly in cases of small tumor regions or low contrast. A key limitation of these methods is that MRI quality degradation is not merely a spatial domain phenomenon but is fundamentally reflected in frequency domain distortion. High-frequency components encode fine details and boundary information, but are particularly susceptible to noise and imaging artifacts in low-quality MRI, while low-frequency components primarily capture global anatomical structures and exhibit higher consistency under different acquisition conditions (see...). Figure 1This observation suggests that traditional spatial domain feature alignment methods are inherently insufficient to handle domain offsets caused by severe quality degradation. Instead, explicit modeling and alignment of frequency domain features provides a more reasonable mechanism for mitigating noise-dominated cross-dataset discrepancies. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an unsupervised cross-domain brain tumor segmentation method for low-quality MRI that can more effectively suppress noise interference and maintain the structural consistency of the tumor region.

[0005] To address the aforementioned technical problems, the present invention provides an unsupervised cross-domain brain tumor segmentation method for low-quality MRI, comprising CDGM-Net. CDGM-Net includes three cooperating components: an initial mask generation module (IMG), a frequency decomposition matching module (FDM), and a frequency cross-fusion module (FCM). Specifically, it includes the following steps:

[0006] Step S1. MRI Image Input and Feature Extraction:

[0007] Input Support Images Supports image masks and query images and will support images With query image The input is fed into a feature extraction network for encoding; the supporting image feature map :

[0008]

[0009] The query image feature map :

[0010]

[0011] in, Represents the feature encoding function;

[0012] Step S2. Initial mask generation:

[0013] First, input the supporting image feature map. Supports image masks and query image feature map Supporting images Perform superpixel segmentation to divide the image into multiple regions with consistent local structure:

[0014]

[0015] in, Indicates the first One superpixel region Indicates the number of regions;

[0016] Subsequently, in support of image masks Under the constraints, a weighted average of the features within each superpixel region is performed to obtain the foreground feature prototype. :

[0017]

[0018] Computational foreground feature prototype With query image features Similarity between them:

[0019]

[0020] Obtain the initial foreground mask for the query image features. ;

[0021] Step S3. Frequency decomposition, i.e., the first stage of the FDM module:

[0022] Input supports image features Query image features The input features are frequency-decomposed using wavelet transform, splitting the feature representation into low-frequency and high-frequency components. The supported image feature decomposition is as follows:

[0023]

[0024] The query image features are decomposed as follows:

[0025]

[0026] in, Representing low-frequency features, used to represent the overall structural information of an image. Representing high-frequency features, used to represent boundary and detail information. This represents support for low-frequency features of images. This represents support for high-frequency features of images, Represents low-frequency features of the query image and Represents high-frequency features of the queried image;

[0027] Step S4. Frequency distribution matching, i.e., the second stage of the FDM module:

[0028] First, input the low-frequency features of the supporting image. Supports high-frequency features of images Query low-frequency features of images and query image high frequency features This will support low-frequency features of images. and query image low-frequency features Normalization to probability distribution and It will support high-frequency features of images. and query image high frequency features Normalization to probability distribution and :

[0029]

[0030] in, and These represent the probability distributions of low-frequency features in the supporting image and the query image, respectively. and These represent the probability distributions of high-frequency features of the supporting image and the query image, respectively.

[0031] Then construct the feature distance matrix. :

[0032]

[0033] in, Indicates the first in the support image feature distribution 1 eigenvector Represents the first element in the feature distribution of the query image. 1 eigenvector;

[0034] The EMD between the two is calculated by solving the optimal transmission problem:

[0035]

[0036] in, Represents the quality transfer matrix;

[0037] By minimizing EMD, frequency distribution alignment between the support image and the query image is achieved, thereby obtaining the low-frequency features of the aligned query image. Aligned query high-frequency image features ;

[0038] Step S5. Frequency cross-fusion, i.e., FCM module:

[0039] Low-frequency features after input alignment Aligned high-frequency features A cross-attention mechanism is used for multi-frequency information fusion, first performing linear mapping. in, , , These represent the query vector, key vector, and value vector, respectively. From the aligned low-frequency features Obtained through linear mapping, it is used to represent the global structural information that needs to be focused on at present; key vector. AND value vector From aligned high-frequency features Obtained through linear mapping;

[0040] Calculate attention weight A:

[0041]

[0042] Based on attention weight fusion features:

[0043]

[0044] Finally, the fused multi-frequency characteristics are obtained:

[0045]

[0046] Step S6. Generate the final segmentation result:

[0047] Input fused multi-frequency features First, global average pooling is performed on the fused features:

[0048]

[0049] in, This indicates a global average pooling operation. A global semantic representation of the fused features.

[0050] Then calculate the foreground probability:

[0051]

[0052] in, For bias parameters, The Sigmoid function is used to calculate the foreground probability. Generate the final brain tumor segmentation mask .

[0053] In step S2, the SLIC algorithm is used to perform superpixel segmentation on the supporting image.

[0054] The initial foreground mask in step S2 This information is input as prior structural information to step S3.

[0055] Advantages of this invention:

[0056] A Frequency-Aware Cross-Dataset Matching Network (CDGM-Net) is proposed to improve the generalization ability of brain tumor segmentation models across low-quality MRI images. By introducing an Initial Mask Generation (IMG) module, a Frequency Decomposition Matching (FDM) module, and a Frequency Cross-Fusion (FCM) module, CDGM-Net systematically alleviates the domain shift problem caused by differences in imaging quality and scanning conditions from a frequency domain perspective, achieving significant improvements in both segmentation accuracy and boundary prediction stability. Extensive experimental results demonstrate that this method performs excellently in cross-dataset transfer tasks from BraTS 2021 to BraTS Africa and BraTS 2020 to BraTSAfrica, especially exhibiting good cross-domain robustness and generalization ability under small sample settings. Compared with traditional single-domain models and existing cross-domain small sample methods, CDGM-Net can more effectively suppress noise interference and maintain the structural consistency of tumor regions in low-quality MRI scenes, validating the effectiveness and practical value of frequency domain alignment and cross-frequency fusion strategies in cross-dataset brain tumor segmentation tasks. Attached Figure Description

[0057] Figure 1 This is a comparative analysis diagram of the frequency domain characteristics of MRI images of different qualities in this invention;

[0058] Figure 2 This is a schematic diagram of the model structure in this invention;

[0059] Figure 3 This is a visualization of the FDM in this invention;

[0060] Figure 4 This is a visual comparison of the segmentation results of each model in the BraTS 2021→BraTS Africa transfer task in this invention;

[0061] Figure 5 This invention provides a comparison of masks with and without SLIC in the IMG.

[0062] Figure 6 This invention visualizes the alignment effect of the FDM module on cross-domain image frequency domain features.

[0063] Figure 7 This is a flowchart of the algorithm of the present invention. Detailed Implementation

[0064] The unsupervised cross-domain brain tumor segmentation method for low-quality MRI of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0065] Example:

[0066] Including CDGM-Net, which consists of three cooperating components: the Initial Mask Generation Module (IMG), the Frequency Decomposition and Matching Module (FDM), and the Frequency Cross-Merging Module (FCM); specifically including the following steps:

[0067] Step S1. MRI Image Input and Feature Extraction:

[0068] Input Support Images Supports image masks and query images and will support images With query image The input is fed into a feature extraction network for encoding; the supporting image feature map :

[0069]

[0070] The query image feature map :

[0071]

[0072] in, Represents the feature encoding function;

[0073] Step S2. Initial mask generation:

[0074] First, input the supporting image feature map. Supports image masks and query image feature map Supporting images Superpixel segmentation is performed using the SLIC algorithm to divide the supporting image into multiple regions with consistent local structures.

[0075]

[0076] in, Indicates the first One superpixel region Indicates the number of regions;

[0077] Subsequently, in support of image masks Under the constraints, a weighted average of the features within each superpixel region is performed to obtain the foreground feature prototype. :

[0078]

[0079] Computational foreground feature prototype With query image features Similarity between them:

[0080]

[0081] Obtain the initial foreground mask for the query image features. , initial foreground mask This is input as prior structural information into step S3;

[0082] Step S3. Frequency decomposition, i.e., the first stage of the FDM module:

[0083] Input supports image features Query image features The input features are frequency-decomposed using wavelet transform, splitting the feature representation into low-frequency and high-frequency components. The supported image feature decomposition is as follows:

[0084]

[0085] The query image features are decomposed as follows:

[0086]

[0087] in, Representing low-frequency features, used to represent the overall structural information of an image. Representing high-frequency features, used to represent boundary and detail information. This represents support for low-frequency features of images. This represents support for high-frequency features of images, Represents low-frequency features of the query image and Represents high-frequency features of the queried image;

[0088] Step S4. Frequency distribution matching, i.e., the second stage of the FDM module:

[0089] First, input the low-frequency features of the supporting image. Supports high-frequency features of images Query low-frequency features of images and query image high frequency features This will support low-frequency features of images. and query image low-frequency features Normalization to probability distribution and It will support high-frequency features of images. and query image high frequency features Normalization to probability distribution and :

[0090]

[0091] in, and These represent the probability distributions of low-frequency features in the supporting image and the query image, respectively. and These represent the probability distributions of high-frequency features of the supporting image and the query image, respectively.

[0092] Then construct the feature distance matrix. :

[0093]

[0094] in, Indicates the first in the support image feature distribution 1 eigenvector Represents the first element in the feature distribution of the query image. 1 eigenvector;

[0095] The Earth Mover's Distance (EMD) between the two is calculated by solving the optimal transport problem:

[0096]

[0097] in, Represents the quality transfer matrix;

[0098] By minimizing EMD, frequency distribution alignment between the support image and the query image is achieved, thereby obtaining the low-frequency features of the aligned query image. and subsequent query high-frequency image features ;

[0099] Step S5. Frequency cross-fusion, i.e., FCM module:

[0100] Low-frequency features after input alignment Aligned high-frequency features A cross-attention mechanism is used for multi-frequency information fusion, first performing linear mapping. in, , , These represent the query vector, key vector, and value vector, respectively. Query vector From the aligned low-frequency characteristics Obtained through linear mapping, it is used to represent the global structural information that needs to be focused on at present; key vector. AND value vector From aligned high-frequency features Obtained through linear mapping.

[0101] Calculate attention weight A:

[0102]

[0103] Based on attention weight fusion features:

[0104]

[0105] Finally, the fused features are obtained as the fused multi-frequency features:

[0106]

[0107] Step S6. Generate the final segmentation result:

[0108] Input fused multi-frequency features First, global average pooling is performed on the fused features:

[0109]

[0110] in, This indicates a global average pooling operation. A global semantic representation of the fused features.

[0111] Then calculate the foreground probability:

[0112]

[0113] in, For bias parameters, The sigmoid function is used. The foreground probability is calculated based on... Generate the final brain tumor segmentation mask .

[0114] experiment:

[0115] 1. Experimental Data: To verify the robustness of CDGM-Net in brain tumor segmentation across datasets, we conducted experiments on three publicly available datasets: BraTS2021, BraTS2020, and BraTS Africa. These datasets exhibit significant differences in image quality, resolution, noise level, and contrast distribution, providing a reasonable testing scenario for the model's generalization ability across MRI images of varying qualities. Descriptions of the three datasets are shown in Table 1.

[0116] Table 1. Core Information of the Dataset

[0117]

[0118] As shown in Table 1, although there are differences in the label definitions of tumor regions between BraTS 2021 / 2020 and BraTS Africa (BraTS 2021 / 2020 uses ET, ED, NCR, while BraTS Africa uses ET, Oedema, NC), most studies use the following three categories as unified segmentation metrics: ET (Enhancing Tumor), TC (Tumor Core), and WT (Whole Tumor). The main reason for this unification is that although the label sets of different datasets differ slightly, their medical meanings are basically the same. A unified segmentation method can reduce labeling bias between datasets and improve the model's cross-dataset generalization ability. The unified mapping definition of tumor partitions for each dataset is shown in Table 2 (ET / TC / WT), and subsequent results and visualizations follow this definition.

[0119] Table 2. Definitions of ET, TC, and WT across different datasets

[0120]

[0121] 2. Implementation details:

[0122] All experiments strictly adhered to a few-shot learning model setup based on source domain training without target domain adaptation. Model parameters were learned entirely from the source domain and remained unchanged during evaluation; inference used only a small number of labeled support samples (1 or 5 samples) from the target domain, without any form of target domain retraining or fine-tuning. This experiment was conducted on an NVIDIA RTX 3090 GPU with 24GB of VRAM. Input data underwent a series of preprocessing operations, including scaling, affine transformation, cropping, RGB dithering, and Gaussian blur, to enhance the model's robustness and generalization ability. During training, 200 images from BraTS 2021 were selected, and one target domain image was introduced to ensure compliance with the few-shot learning setting. Model optimization employed stochastic gradient descent (SGD) with a batch size of 1, an initial learning rate of 0.001, and a total of 100 training epochs. Furthermore, to mitigate overfitting and ensure model robustness, a 5-fold cross-validation strategy was used for evaluation.

[0123] The AdamW optimizer (with weight decay) and cosine annealing learning rate were used. A brief linear warm-up was performed before training, and the scheduling parameters were consistent throughout the paper. The random seed was fixed at 42 to ensure that PyTorch, NumPy, and CUDA used the same random seed, and deterministic operators were enabled. The results in this paper are from a single training and inference run. The relevant training logs, configurations, and model weights can be found in the supplementary materials. The model was selected based on the highest checkpoint on the validation set Dice, and then evaluated once on the test set. The test set was not involved in any form of hyperparameter tuning.

[0124] To quantitatively evaluate segmentation performance after each training epoch, this paper uses two metrics: the Dice similarity coefficient (DSC) and the 95% Hausdorff Distance (HD95).

[0125]

[0126] in, and These represent the predicted and actual foreground pixel sets, respectively. DSC measures the degree of overlap between the two, with a value ranging from 0 to 1. The larger the value, the more consistent the segmentation result is with the actual annotation.

[0127]

[0128]

[0129] in, and For the predicted and true boundary point set, The distance is Euclidean distance. HD95 measures the maximum distance between the predicted boundary and the true boundary. The smaller the value, the more accurate the boundary localization.

[0130] 3. Performance Test

[0131] To fully verify the cross-domain generalization ability of CDGM-Net, we conducted a systematic comparative experiment in the transfer scenarios of BraTS 2021→BraTS Africa and BraTS 2020→BraTS Africa. The experiment included two types of baselines: ① single-domain models (U-Net

[39] , V-Net

[39] , MSA-VNet

[39] , nnUNet

[40] ), which reflect the transfer performance of traditional models in cross-dataset environments; ② cross-domain few-shot segmentation models (PATNet

[26] , HDMNet

[41] , RestNet

[42] , ABCDFSS

[43] , TGCM

[44] ), which represent the current state-of-the-art cross-domain few-shot methods. All models were trained on BraTS 2021 and BraTS 2020 and tested on BraTS Africa to evaluate segmentation performance in three tumor regions: ET (tumor extension), TC (tumor core), and WT (whole tumor region). Tables 3 and 4 show the performance comparison of each model in the BraTS 2021→BraTS Africa and BraTS 2020→BraTS Africa transfer scenarios, respectively. Since references [37-39] do not provide transfer experiment data for BraTS 2020→BraTS Africa, this paper only provides experimental data comparisons for the open-source algorithms. Table 4 shows the performance comparison of each model in the BraTS 2020→BraTS Africa transfer scenario. In the table, the best value is marked in bold, and the second best value is underlined.

[0132] Table 3. Performance comparison of each model on BraTS 2021 BraTS Africa

[0133]

[0134] As shown in Table 3, in the BraTS 2021→BraTS Africa cross-dataset transfer task, CDGM-Net achieved more stable and competitive performance in the ET, TC, and WT tumor regions. In particular, it outperformed single-domain models and existing cross-domain small-sample methods in terms of HD95, indicating that it has stronger boundary prediction stability and cross-domain generalization ability in low-quality MRI scenarios. In contrast, traditional single-domain models (U-Net, V-Net, MSA-VNet, and nnUNet) generally showed significant degradation when transferred to BraTS Africa, with a significant increase in HD95 (e.g., nnUNet reached 34.88 mm and 38.24 mm in the ET and TC regions, respectively). This indicates that models relying solely on spatial domain features are unable to cope with the structural uncertainties caused by the combined effects of image quality degradation and domain shift.

[0135] Existing cross-domain few-shot methods (such as PATNet, HDMNet, RestNet, ABCDFSS, and TGCM) have shown certain advantages over single-domain models in this transfer task, validating the effectiveness of prototype learning, cross-domain transformation, or data augmentation strategies in few-shot scenarios. However, these methods mainly perform alignment or enhancement at the spatial feature level. When faced with the high noise, low contrast, and blurred boundary problems prevalent in BraTS Africa, they still struggle to stably distinguish between real tumor boundaries and noise textures, resulting in an overestimation of HD95 in some regions (especially ET and TC). In contrast, CDGM-Net explicitly aligns the high and low frequency distributions of the source and target domains at the frequency domain level through FDM, effectively mitigating the spectral shift caused by imaging style differences. Furthermore, by combining the structurally consistent foreground prior provided by IMG with the cross-frequency fusion mechanism of FCM, fragmented prediction and boundary spillover phenomena are further suppressed. Therefore, CDGM-Net can achieve more consistent boundary localization performance while maintaining DSC comparable to or even higher than state-of-the-art cross-domain methods, demonstrating its unique advantages in low-quality MRI cross-dataset segmentation tasks.

[0136] Table 4. Performance Comparison of Each Model in BraTS 2020 and BraTS Africa

[0137]

[0138] As shown in Table 4, in the BraTS 2020 → BraTS Africa cross-dataset transfer task, various cross-domain few-shot methods generally performed more stably than single-domain models, indicating that prototype learning, cross-domain transformation, and data augmentation strategies are effective in low-shot transfer scenarios. Among them, PATNet (ResNet) and TGCM achieved higher DSC in the 5-shot setting and outperformed some methods in boundary metrics, indicating that they mitigated the performance degradation caused by domain shift to some extent. However, these methods mainly rely on spatial domain feature alignment or sample-level augmentation. When faced with the high noise, low contrast, and boundary blurring problems prevalent in BraTS Africa, their HD95 remains relatively high, especially in structurally complex regions such as ET and TC, where boundary localization stability is still limited.

[0139] In contrast, CDGM-Net exhibits more consistent performance advantages in this migration scenario. In the 1-shot setting, it maintains relatively balanced DSC and HD95 across the ET, TC, and WT regions, demonstrating good robustness in small-sample scenarios. In the 5-shot setting, CDGM-Net's DSC further improves to 84.50%, 83.46%, and 84.13%, while HD95 significantly decreases to 26.87mm, 23.12mm, and 24.98mm, outperforming existing cross-domain small-sample methods overall. This advantage is mainly attributed to the introduction of a frequency domain modeling mechanism: FDM explicitly aligns the high and low frequency distributions of the source and target domains at the spectral level, fundamentally mitigating the spectral shift caused by differences in imaging quality; simultaneously, the structurally consistent foreground prior provided by IMG and the cross-frequency fusion mechanism of FCM effectively suppress noise-dominated fragmented predictions and boundary spillover problems. Therefore, CDGM-Net not only maintains its leading position in segmentation accuracy, but also demonstrates a significant advantage in boundary prediction stability, verifying its robustness and generalization ability in low-quality MRI cross-dataset segmentation tasks.

[0140] To more intuitively demonstrate the segmentation performance of the proposed method, we selected several typical models for comparison, showcasing their segmentation results in the BraTS 2021→BraTS Africa transfer task. Specifically, we chose 5-shot results from cross-domain few-shot methods for comparison. Figure 4 The segmentation performance of each model in this task is demonstrated. Figure 4Observations show that in the BraTS 2021→BraTS Africa cross-domain transfer task, CDGM-Net demonstrated significant advantages in segmenting the ET, TC, and WT tumor regions. CDGM-Net accurately segmented the tumor region with clear boundaries and high consistency with the Ground Truth. In the ET region, CDGM-Net performed excellently. While other models such as TGCM and ABCDGSS performed well, their segmentation details were somewhat blurry. PATNet (ResNet), nnUNet, MSA-VNet, and U-Net performed poorly on low-quality images, with blurred and incomplete tumor boundaries. In the TC region, CDGM-Net continued to lead with accurate segmentation results, especially on low-quality MRI images. In comparison, TGCM and ABCDGSS had slightly lower segmentation accuracy, PATNet (ResNet) and nnUNet had somewhat blurry boundary localization, and MSA-VNet and V-Net performed worse than other methods. For the WT region, CDGM-Net successfully identified the tumor extension, core, and other regions, especially excelling in the segmentation of small targets and complex tumor structures. TGCM and ABCDGSS performed well, but their boundary accuracy was still inferior to CDGM-Net. PATNet (ResNet) and nnUNet showed weaker segmentation results. In summary, CDGM-Net demonstrated superior segmentation accuracy and stability in all tumor regions, especially in low-quality MRI images and small sample settings, where its performance far surpassed other models. By introducing frequency domain processing and cross-domain attention mechanisms, CDGM-Net effectively reduced noise interference and maintained the structural consistency of tumor regions, demonstrating excellent cross-dataset transferability and robustness.

[0141] 4. Computational efficiency:

[0142] To evaluate the computational efficiency of the proposed CDGM-Net in practical applications and to verify whether it introduces additional computational burden while improving performance, this paper systematically compares it with three representative segmentation models—U-Net, V-Net, and nnU-Net—on multiple computational efficiency metrics. Specifically, the methods were quantitatively evaluated from four aspects: model parameter count (params), floating-point operations (FLOPs), inference latency (latency), and memory usage (memory). All experiments were conducted on a cross-dataset setup of BraTS 2021→BraTS Africa and tested in the same hardware and software environment. To ensure fairness in the comparison, all models used ResNet-50 as the backbone network, and the input image size was uniformly set to 400×400. The relevant computational efficiency comparison results are summarized in Table 5.

[0143] Table 5 Comparison of computational efficiency

[0144]

[0145] Table 5 presents a detailed comparison of the computational complexity and resource consumption of different methods. It can be observed that CDGM-Net has 34.2M parameters, which is on the same order of magnitude as U-Net (31.8M) and nnU-Net (33.5M), indicating that the introduction of IMG, FDM, and FCM modules did not significantly increase the model size. Meanwhile, CDGM-Net's FLOPs are 73.8G, only slightly higher than U-Net and nnU-Net, suggesting that the additional computational overhead from frequency domain modeling and cross-frequency fusion is limited and controllable.

[0146] In terms of inference efficiency, CDGM-Net's average inference latency is 44.7ms, significantly lower than nnU-Net's 63.4ms, and close to U-Net's (42.1ms), demonstrating a superior inference speed. Furthermore, its memory usage is 3.5GB, comparable to V-Net and nnU-Net, without significant memory expansion. This result indicates that although CDGM-Net introduces frequency domain decomposition matching and cross-attention fusion mechanisms into its network structure, these modules primarily operate on low-to-medium resolution feature layers, without significantly impacting overall inference efficiency or memory requirements.

[0147] In summary, the above analysis leads to the conclusion that CDGM-Net significantly improves cross-dataset segmentation performance while maintaining computational complexity and inference efficiency comparable to mainstream segmentation models. This method achieves a good balance between segmentation accuracy, boundary stability, and computational efficiency, making it feasible for deployment in real-world clinical environments, and particularly suitable for medical image segmentation applications with high requirements for inference speed and hardware resources.

[0148] 5. Ablation test

[0149] To verify the impact of each module on model performance, ablation experiments were conducted on the backbone, IMG, FDM, and FCM modules. First, the impact of different backbones on the overall model performance was compared. Second, the superpixel clustering (SLIC) in the IMG module was analyzed. Finally, the impact of the frequency domain matching mechanism of FDM and the fusion mechanism of FCM on the final segmentation results was analyzed.

[0150] (1) Backbone ablation experiment

[0151] To systematically evaluate the impact of different feature extraction networks (Backbone) on the overall segmentation performance of CDGM-Net, we designed a backbone network ablation experiment. This experiment compared three representative pre-trained backbone networks: VGG16, ResNet50, and DINOv2, representing classic shallow convolutional networks, deep residual networks, and the self-supervised pre-trained Vision Transformer architecture, respectively. While keeping other modules and training strategies completely consistent, the experiment only replaced the feature extraction backbone and compared performance on a cross-dataset segmentation task from BraTS 2021 to BraTS Africa; Table 5 shows the experimental results under different backbone networks.

[0152] Table 6. Performance Comparison of Different Backbone Networks in Cross-Domain Segmentation Tasks

[0153]

[0154] As shown in Table 6, different backbone networks significantly impact the performance of CDGM-Net in cross-dataset segmentation tasks. ResNet50 achieved optimal performance in both 1-shot and 5-shot settings. Particularly in the 5-shot setting, its DSC (Digital Subtraction Cross-Segmentation) in the ET, TC, and WT tumor regions reached 86.52%, 84.86%, and 87.22%, respectively, while the corresponding HD95 values ​​significantly decreased to 18.85mm, 19.59mm, and 19.70mm, indicating that this model excels in both segmentation accuracy and boundary stability.

[0155] The superior performance of ResNet50 is primarily attributed to the deep feature modeling advantages brought about by its residual connections. By introducing identity mappings, the residual structure effectively alleviates the gradient vanishing and feature degradation problems during deep network training, enabling the network to learn richer high-level semantic features while maintaining stable optimization. This characteristic is particularly crucial for brain tumor segmentation, especially in low-quality MRI images where tumor boundaries are often blurred and tissue contrast is low, requiring the model to rely on multi-level features to gradually aggregate local details and global structural information. ResNet50 can simultaneously preserve edge cues and structural context in deep features, thereby significantly improving the localization accuracy and boundary consistency of tumor regions.

[0156] In contrast, VGG16's performance in the 1-shot setting is significantly weaker, with DSC values ​​of only 79.54%, 78.90%, and 79.12% for the ET, TC, and WT regions, respectively, while its HD95 value is relatively high. This phenomenon is mainly due to the lack of an explicit cross-layer information transfer mechanism in the pure stacked convolutional structure used by VGG16, which easily leads to limited feature representation capabilities as the network deepens. Especially in low-quality MRI images, noise and uneven contrast further amplify the instability of shallow features, making it difficult for the model to effectively capture fine-grained boundary information of complex tumor structures.

[0157] DINOv2, a Vision Transformer architecture based on self-supervised pre-training, falls between ResNet50 and VGG16 in overall performance. Its DSC performance in a 1-shot setting is superior to VGG16, indicating that self-supervised pre-training helps improve feature generalization ability. In a 5-shot setting, its performance further improves and approaches ResNet50 in the Wt region. However, compared to convolutional networks, the Vision Transformer still has certain limitations in modeling local texture and edge information under small sample conditions, especially in low-quality MRI scenes, where local detail noise may interfere with global attention modeling, thus affecting boundary accuracy.

[0158] Based on the above analysis, ResNet50 exhibits the best overall fit within the CDGM-Net framework. Its residual structure not only enhances the expressive power of deep features but also improves the model's training stability and segmentation robustness in low-quality MRI images and cross-dataset scenarios. Therefore, ResNet50 was selected as the default backbone network in subsequent experiments.

[0159] (2) Ablation experiment of IMG

[0160] To further verify the role of the Initial Mask Generation Module (IMG) in cross-dataset segmentation tasks, this paper conducts an ablation comparison experiment with and without the IMG module. In this experiment, the Backbone, FDM, FCM, and training strategies were kept completely consistent; only the IMG module was removed, and the original features were directly used in the subsequent frequency domain matching and fusion process. The experimental results are shown in Table 7.

[0161] Table 7 Ablation experiments of IMG

[0162]

[0163] Table 7 shows that the introduction of the IMG module resulted in a stable improvement in segmentation performance across the ET, TC, and WT tumor regions. Particularly in the 1-shot setting, DSC improved by approximately 1–2% compared to the configuration without IMG, while HD95 decreased by an average of approximately 1–3 mm across all regions, indicating that the IMG module has a more significant positive impact on segmentation accuracy and boundary stability under extremely small sample conditions. From the perspective of different tumor regions, the performance gain brought by the IMG module was most significant in the ET region. This is mainly because the ET region typically presents as a small-volume, poorly defined lesion structure, highly sensitive to noise and contrast changes. Through the structured region segmentation generated by superpixel clustering, the IMG module can effectively suppress random noise interference in the early stages of feature extraction, making the foreground prototype more concentrated in the real lesion region, thus providing a more stable structural prior for subsequent feature matching. Furthermore, in the 5-shot setting, although the overall performance improvement was slightly smaller than in the 1-shot setting, the IMG module still consistently improved DSC and reduced HD95 across all tumor regions. This indicates that IMG is not only effective in extremely small sample scenarios, but also enhances the model's generalization stability across different datasets by improving the quality of feature distribution, even with an increased sample size. In summary, the IMG module effectively improves the segmentation performance of CDGM-Net across various sample sizes by introducing a noisy, robust, and structurally consistent initial foreground representation at the model's front end. Its performance gain does not come from additional supervision information, but rather from improving feature quality, providing a more reliable input foundation for subsequent frequency domain decomposition matching and cross-frequency fusion.

[0164] Taking BraTS Africa case 10 as an example ( Figure 5 Compared to the gold standard, the IMG+SLIC mask (right) performs best in terms of tumor boundary continuity and geometric smoothness; while the IMG w / o SLIC (middle) exhibits noise-induced local boundary irregularities (jagged edges / burrs) and internal false negative holes (marked by red arrows). This comparison directly verifies that superpixel-based SLIC region consistency constraints can effectively suppress fragmented false detections, improve the morphological fidelity and cross-domain robustness of the initial mask, and thus provide more reliable input for subsequent alignment and fusion processes.

[0165] (3) Ablation experiment of FDM module

[0166] To verify the effectiveness of the Frequency Domain Decomposition Matching (FDM) module in cross-dataset brain tumor segmentation, this paper further designed an ablation experiment targeting FDM. In this experiment, the Backbone, IMG, FDM, and training strategy were kept unchanged, and only the FDM module was removed. That is, frequency domain decomposition and EMD alignment of high and low frequency features were not performed; instead, the original features were directly fed into the subsequent fusion and prediction stages. The experimental results are as follows: Figure 6As shown.

[0167] from Figure 6 The content shown can be analyzed to show that:

[0168] ① Low-frequency feature alignment: Low-frequency features of the processed query image ( Histogram of distribution of ) Figure 6 (third row, first column) and supporting images ( The results are closer, indicating that FDM effectively corrects the global style shift caused by differences in brightness and contrast;

[0169] ② High-frequency feature enhancement and alignment: High-frequency features of the processed query image ( The distribution of sharpness values ​​() Figure 6 The image in the third row and second column is located between the original query image and the supporting image, and the difference heatmap shows that the difference between it and the supporting image has decreased, indicating that FDM enhances the blurred edges while suppressing high-frequency noise and achieving domain adaptation at the level of detail.

[0170] In summary, numerical enhancement, controllable computational cost, and visual analysis of feature distribution together constitute a solid chain of evidence demonstrating that our proposed frequency domain matching and fusion mechanism can systematically solve the domain generalization problem caused by uneven image quality.

[0171] (4) FCM ablation experiment

[0172] To further verify the effectiveness of the cross-attention mechanism in the Frequency Domain Cross-Fusing Module (FCM), this paper designs ablation experiments for different frequency domain feature fusion strategies. In this experiment, the Backbone, IMG, and FDM modules are kept unchanged, and only the fusion method of high- and low-frequency features is changed. Simple addition, channel concatenation, and cross-attention fusion are used for comparative analysis. This ablation experiment aims to verify the advantages of the cross-attention mechanism over simple fusion strategies in coordinating structural information and boundary details. The experimental results are shown in Table 8.

[0173] Table 8. Performance comparison of different frequency domain feature fusion strategies on the BraTS 2021→BraTS Africa cross-domain segmentation task.

[0174]

[0175] Table 8 shows that different frequency domain feature fusion strategies have a significant impact on the segmentation performance of brain tumors across datasets. Among them, the fusion method based on cross-attention mechanism (Cross-Attention) achieved the best results in both 1-shot and 5-shot settings, verifying the effectiveness of the FCM module in multi-frequency feature collaborative modeling. In the 1-shot setting, Cross-Attention achieved DSCs of 83.50%, 81.33%, and 83.17% in the ET, TC, and WT regions, respectively, showing a stable advantage over the Add and Concat strategies. Simultaneously, its HD95 decreased to 23.20 mm, 24.14 mm, and 24.06 mm, respectively, indicating that cross-attention fusion has higher stability in boundary prediction. In the 5-shot setting, although the overall performance further improved, Cross-Attention still maintained the best performance in each tumor region, with DSCs reaching 86.52%, 84.86%, and 87.22%, respectively, and HD95 further decreasing to the 18.85–19.70 mm range. From the perspective of different tumor regions, the FCM module shows more significant performance improvements in the ET and WT regions. This is mainly because these two types of regions are highly sensitive to structural consistency and boundary details. Simple feature addition or splicing is difficult to effectively coordinate high- and low-frequency information, while the cross-attention mechanism can explicitly model the dependencies between different frequency bands, thereby achieving more consistent feature fusion. In summary, compared with static fusion methods, the cross-attention-based FCM module can adaptively balance structural information and detailed features, significantly improving segmentation accuracy and boundary stability under cross-dataset and small sample conditions. These experimental results are consistent with the design motivation of FCM and further verify the necessity of the cross-attention mechanism in low-quality MRI cross-domain segmentation tasks.

[0176] (5) Ablation experiments of each component of the model

[0177] To comprehensively evaluate the contribution of each module to the overall model performance, we designed an overall ablation experiment. This experiment used ResNet50 as the backbone network, progressively enabling the IMG, FDM, and FCM modules, and evaluated performance in a 5-shot setting. We quantified the overall impact of each module on the model by comparing the segmentation accuracy (DSC and HD95) and computational cost (number of parameters and inference latency) under different module combinations. The experimental results are shown in Table 9.

[0178] Table 9 Ablation experiments of each component

[0179]

[0180] As shown in Table 9, the gradual introduction of each module significantly improved the segmentation accuracy and boundary accuracy of the model. Enabling the IMG module significantly improved the model's segmentation accuracy (DSC). Although HD95 improved, using the IMG module alone did not significantly improve boundary accuracy, indicating that the IMG module mainly improved the stability of the foreground region, while boundary accuracy still had room for improvement. Introducing the FDM module further improved the model's performance, with DSC reaching 84.82%, 83.24%, and 83.55%, respectively, and HD95 significantly reduced. The FDM module reduced the domain offset between low-quality MRI and high-quality images through frequency domain alignment, thereby improving the stability and accuracy of cross-dataset segmentation. When the FCM module was added, the performance reached its optimal level. DSC further improved to 86.52%, 84.86%, and 87.22%, while HD95 decreased to 18.85mm, 19.59mm, and 19.70mm. FCM enhanced the fusion of low-frequency and high-frequency features through a cross-domain attention mechanism, improving the model's segmentation accuracy and boundary stability for complex tumor structures. Although the number of model parameters and inference latency increased with the addition of modules (ultimately reaching 34.2M and 44.7ms respectively), the increase in computational cost was acceptable compared to the improvement in accuracy. In summary, the gradual introduction of IMG, FDM, and FCM modules significantly improved segmentation accuracy, boundary stability, and generalization ability across datasets, especially in the processing of low-quality MRI images.

[0181] This paper proposes a novel frequency-aware cross-dataset matching network (CDGM-Net) to improve the generalization ability of brain tumor segmentation models across low-quality MRI images. By introducing an Initial Mask Generation (IMG) module, a Frequency Decomposition Matching (FDM) module, and a Frequency Cross-fusion (FCM) module, CDGM-Net systematically alleviates the domain shift problem caused by differences in imaging quality and scanning conditions from a frequency domain perspective, achieving significant improvements in both segmentation accuracy and boundary prediction stability. Extensive experimental results demonstrate that this method performs excellently in cross-dataset transfer tasks from BraTS 2021 to BraTS Africa and BraTS 2020 to BraTS Africa, especially exhibiting good cross-domain robustness and generalization ability under small sample settings. Compared with traditional single-domain models and existing cross-domain small sample methods, CDGM-Net can more effectively suppress noise interference and maintain the structural consistency of tumor regions in low-quality MRI scenes, validating the effectiveness and practical value of frequency domain alignment and cross-frequency fusion strategies in cross-dataset brain tumor segmentation tasks. While CDGM-Net has achieved significant performance improvements in low-quality MRI segmentation tasks, it still has certain limitations under extreme imaging conditions. For example, in scenes with extremely high noise levels or severely degraded contrast, high-frequency components may be dominated by noise, thus affecting the stability of boundary prediction. This indicates that existing frequency domain modeling strategies still have room for further optimization under complex noise distributions. Future research will mainly focus on two directions: first, exploring more refined multi-scale frequency domain feature modeling strategies to enhance the model's ability to express complex tumor structures and weak boundaries; second, introducing noise-aware or adaptive spectral constraint mechanisms within the existing frequency domain framework to further suppress noise-induced high-frequency interference, thereby improving the model's segmentation stability and robustness under extremely low-quality MRI conditions.

Claims

1. An unsupervised cross-domain brain tumor segmentation method for low-quality MRI, characterized in that, This includes CDGM-Net, which comprises three cooperating components: an Initial Mask Generation Module (IMG), a Frequency Decomposition and Matching Module (FDM), and a Frequency Cross-Merging Module (FCM); specifically, it includes the following steps: Step S1. MRI Image Input and Feature Extraction: Input Support Images Supports image masks and query images and will support images With query image The input is fed into a feature extraction network for encoding; the supporting image feature map : ； The query image feature map : ； in, Represents the feature encoding function; Step S2. Initial mask generation: First, input the supporting image feature map. Supports image masks and query image feature map Supporting images Perform superpixel segmentation to divide the image into multiple regions with consistent local structure: ； in, Indicates the first One superpixel region Indicates the number of regions; Subsequently, in support of image masks Under the constraints, a weighted average of the features within each superpixel region is performed to obtain the foreground feature prototype. : ； Computational foreground feature prototype With query image features Similarity between them: ； Obtain the initial foreground mask for the query image features. ; Step S3. Frequency decomposition, i.e., the first stage of the FDM module: Input supports image features Query image features The input features are frequency-decomposed using wavelet transform, splitting the feature representation into low-frequency and high-frequency components. The supported image feature decomposition is as follows: ； The query image features are decomposed as follows: ； in, Representing low-frequency features, used to represent the overall structural information of an image. Representing high-frequency features, used to represent boundary and detail information. This represents support for low-frequency features of images. This represents support for high-frequency features of images, Represents low-frequency features of the query image and Represents high-frequency features of the queried image; Step S4. Frequency distribution matching, i.e., the second stage of the FDM module: First, input the low-frequency features of the supporting image. Supports high-frequency features of images Query low-frequency features of images and query image high frequency features This will support low-frequency features of images. and query image low-frequency features Normalization to probability distribution and It will support high-frequency features of images. and query image high frequency features Normalization to probability distribution and : ； in, and These represent the probability distributions of low-frequency features in the supporting image and the query image, respectively. and These represent the probability distributions of high-frequency features of the supporting image and the query image, respectively. Then construct the feature distance matrix. : ； in, Indicates the first in the support image feature distribution 1 eigenvector Represents the first element in the feature distribution of the query image. 1 eigenvector; The EMD between the two is calculated by solving the optimal transmission problem: ； in, Represents the quality transfer matrix; By minimizing EMD, frequency distribution alignment between the support image and the query image is achieved, thereby obtaining the low-frequency features of the aligned query image. Aligned query high-frequency image features ; Step S5. Frequency cross-fusion, i.e., FCM module: Low-frequency features after input alignment Aligned high-frequency features A cross-attention mechanism is used for multi-frequency information fusion, first performing linear mapping. in, , , These represent the query vector, key vector, and value vector, respectively. From the aligned low-frequency features Obtained through linear mapping, it is used to represent the global structural information that needs to be focused on at present; key vector. AND value vector From aligned high-frequency features Obtained through linear mapping; Calculate attention weight A: ； Based on attention weight fusion features: ； Finally, the fused multi-frequency characteristics are obtained: ； Step S6. Generate the final segmentation result: Input fused multi-frequency features First, global average pooling is performed on the fused features: ； in, This indicates a global average pooling operation. A global semantic representation of the fused features; Then calculate the foreground probability: ； in, For bias parameters, The Sigmoid function is used to calculate the foreground probability. Generate the final brain tumor segmentation mask .

2. The unsupervised cross-domain brain tumor segmentation method for low-quality MRI according to claim 1, characterized in that: In step S2, the SLIC algorithm is used to perform superpixel segmentation on the supporting image.

3. The unsupervised cross-domain brain tumor segmentation method for low-quality MRI according to claim 1, characterized in that: The initial foreground mask in step S2 This information is input as prior structural information into step S3.

Images