EdgeFormer-Med: Boundary-Adaptive Transformer for Precise Segmentation of Small and Low-Contrast Anatomical Structures

Abstract

Accurate segmentation of brain tumors from magnetic resonance imaging (MRI) is an important task in computer-aided diagnosis, treatment planning, and disease monitoring. However, small tumor regions and irregular anatomical structures remain difficult to segment because of weak boundaries, low contrast, class imbalance, scale variation, and the loss of fine spatial information during feature extraction and downsampling. Existing medical image segmentation approaches have explored boundary-aware attention, multi-scale feature learning, Transformer-based global context modeling, and specialized loss functions, but these techniques often address individual aspects of the problem. This paper proposes EdgeFormer-Med, a boundary-adaptive Transformer framework designed for precise segmentation of small and low-contrast anatomical structures in brain MRI. The proposed framework combines multi-scale feature representation with Transformer-based contextual modeling and an edge-guided attention mechanism to preserve fine boundary information. A boundary-adaptive feature fusion stage is proposed to integrate edge, local, and global features, while a small-object-aware loss is designed to place greater optimization emphasis on difficult and small target regions. The framework is planned for evaluation using the BraTS Adult Glioma (BraTS-GLI) benchmark dataset and standard segmentation metrics such as Dice Similarity Coefficient, Intersection over Union, Precision, Recall, and HD95. As this work is presented as a research framework without completed experiments, no numerical performance results are claimed. The proposed methodology establishes a structured basis for future implementation, quantitative evaluation, and ablation analysis.