GSCA-UNet: a gated spatial-channel attention U-net for accurate skin lesion segmentation

Abstract

Medical image segmentation is a fundamental component of computer-aided diagnosis, where automatic skin lesion segmentation serves as a critical upstream task by providing pixel-wise delineations for subsequent analysis. Deep encoder-decoder architectures, such as U-Net and its variants, have advanced skin lesion segmentation. However, the task remains challenging. Key difficulties include low contrast between lesions and skin, ambiguous or irregular boundaries, acquisition artifacts. Furthermore, lesions exhibit large intra-class variations in scale, shape and texture. In this work, we propose GSCA-UNet (Gated Spatial-Channel Attention UNet), a novel segmentation architecture for skin lesions. At its core, a gated spatial attention block adaptively models horizontal and vertical spatial dependencies by multi-scale 1D convolutions with learnable gating to strengthen lesion boundaries while suppressing background clutter. A cross-dimensional attention interaction block establishes bidirectional guidance between spatial and channel attention through multi-head self-attention and gating fusion. Extensive experiments on public benchmark datasets demonstrate that GSCA-UNet consistently outperforms competitive baselines across multiple metrics and exhibits superior robustness on challenging cases with blurry borders, irregular shapes, and severe artifacts.