Enhancing Brain Tumor Segmentation with Attention-Based Feature Fusion in Swin-UNETR

Brain tumor segmentation remains a major challenge in medical image analysis, due to its crucial importance in disease diagnosis and monitoring. However, variations in tumor shape, size, and intensity across different MRI modalities continue to pose significant challenges for automated segmentation methods. Therefore we are exploring the combination of the most effective deep learning techniques, from CNNs to Transformers in a U-shaped architecture. We present A-SWIN-UNETR, an improved version of the baseline SWIN-UNETR, in which, instead of the basic concatenation between the Transformer encoder and the CNN decoder, we use a spatial and channel attention mechanisms to capture both local and global features by focusing on tumor boundaries for more accurate segmentation. Our approach achieves superior performance on the BraTS-Lighthouse challenge dataset, with the channel attention variant showing particularly strong results with an average Dice score of 0.902 and substantial improvements in Hausdorff distance (HD95) metrics. The proposed attention mechanisms allow the model to learn both WHERE to focus spatially and WHICH features to extract, leading to more efficient feature fusion and improved segmentation accuracy.