A Boundary-Guided Attention U-Net for Improved Brain Tumor Segmentation from MRI Images

Accurate brain tumor segmentation from magnetic resonance imaging (MRI) plays an important role in diagnosis and treatment planning. However, existing deep learning models, such as U-Net and its variants, often struggle to clearly capture tumor boundaries, especially in low-contrast and complex regions. In this work, we propose a Boundary-Guided Attention U-Net (BGA-U-Net), which enhances the standard U-Net by combining attention mechanisms with a dedicated boundary prediction branch. This allows the model to better focus on relevant tumor regions while improving edge detection. We evaluate the proposed method on the FLAIR modality of the BraTS 2020 dataset and compare it with U-Net and Attention U-Net. The results show that BGA-U-Net achieves a Dice score of 0.8363, along with high specificity (0.997) and strong sensitivity. To further validate the proposed method, we conduct a Wilcoxon signed-rank statistical significance test (p = 0.0312), a comprehensive ablation study, and boundary-specific evaluations using HD95 and ASD metrics, all confirming the reliability and effectiveness of the proposed approach. Overall, the model provides more accurate and reliable segmentation, highlighting the benefit of boundary-aware learning.

Hali tarjima qilinmagan