Skin Lesion Segmentation Based on a Boundary-Aware Multi-Scale Attention Network

Accurate segmentation of skin lesions from dermoscopic images is fundamentally challenged by high intra-class variance, ill-defined boundaries, and imaging artifacts. Existing convolutional and hybrid CNN–Transformer models often yield coarse masks that miss fine boundary details. To address this issue, we propose GLR-Net, a hierarchical encoder–decoder architecture that leverages multi-level attention to jointly model global, local, and region-level cues. At its core, GLR-Net integrates four synergistic components—Global–Local Refinement (GLoR), Boundary–Semantic Integration and Selection Filter (BISF), Attention-Guided Context Refinement (AGCR), and multi-scale Reverse Attention (RA)—so that deep encoder stages are first refined to be boundary-aware and decoder stages are subsequently sharpened in a top–down manner. Evaluated on four public dermoscopic benchmarks (ISIC 2016, ISIC 2017, ISIC 2018, and PH2), GLR-Net attains competitive or superior performance under the official ISIC protocols. In particular, on the challenging ISIC 2017 dataset, it achieves 92.32% Dice and 90.71% Intersection over Union (IoU), surpassing strong open-source baselines such as BiFBA-Net while remaining competitive with more recent models such as MRP-UNet. These results, together with consistently high accuracy across datasets, confirm that an explicit multi-scale, boundary-centric design is effective for improving boundary fidelity in dermoscopic lesion segmentation.