Edge-enhanced 3D attention U-Net for accurate brain tumor delineation and classification in magnetic resonance imaging

Recent advances in biomedical imaging and deep learning have enabled new approaches for noninvasive tumor analysis and intraoperative guidance. This study proposes a two-stage deep learning framework for brain tumor segmentation and classification from multi-modal MRI, with potential integration into optical-image-guided diagnostic systems. The first stage performs volumetric tumor segmentation using an enhanced 3D U-Net equipped with a novel Edge-Enhanced Convolutional Block Attention Module (EE-CBAM), which combines Sobel-based edge extraction with spatial and channel attention to refine boundary precision and improve spatial focus. This mechanism reinforces the model’s sensitivity to fine tumor structures and preserves anatomical continuity, which are critical for accurate interpretation in optical and MRI fusion-based workflows. In the second stage, the segmented tumor region is classified into low-grade gliomas (LGG) or high-grade gliomas (HGG) using a 3D CNN optimized with class-weighted loss to mitigate dataset imbalance. Experiments on the MICCAI BraTS dataset demonstrate that the proposed model achieves a Dice coefficient of 0.81 and an Intersection over Union (IoU) of 0.75 for segmentation, while the classification network attains F1-scores of 0.74 (LGG) and 0.92 (HGG). These results confirm the framework’s strong generalization capability and clinical relevance. The proposed EEA-U-Net offers a robust AI-assisted pipeline for precise tumor delineation and grading, providing a promising foundation for real-time integration into optical imaging or intraoperative navigation systems.