BrainAdaptNet: A few-shot learning model for brain tumor segmentation in low-quality MRI

Segmentation of brain tumors in low-quality magnetic resonance imaging (MRI) remains a challenge, especially in resource-limited regions such as sub-Saharan Africa. Variations in MRI data quality lead to poor model generalization, and low-quality data is difficult to acquire and annotate. To address these issues, this article proposes a few-shot cross-domain segmentation model, BrainAdaptNet, aimed at improving segmentation performance from high-quality to low-quality MRI. BrainAdaptNet uses DinoV2 as a self-supervised feature extraction backbone, enhancing dataset adaptability through contrastive learning and simulating image quality degradation with wavelet transform style transfer to improve the model’s robustness to changes in image distribution. In the data preprocessing phase, CutMix augmentation is introduced for feature fusion, reducing the distribution gap between source and target data. Additionally, a task-adaptive cross-attention (TACA) feature transformation module is designed to further enhance the interaction and alignment of features between the support and query sets, optimizing segmentation performance under few-shot conditions. To rigorously evaluate the proposed method, fivefold cross-validation was conducted on data migration tasks with BraTS 2021 and BraTS 2020 as the source domains and BraTS Africa as the target domain. Experimental results show that, compared to the baseline model PATNet, BrainAdaptNet achieved a 1.5–2.0 percentage point improvement in the dice similarity coefficient (DSC) and a 4.8–5.2 percentage point reduction in Hausdorff distance (HD95) in cross-dataset segmentation tasks, demonstrating the effectiveness of BrainAdaptNet on low-quality MRI data.