RGBA-UNet: An Ultra-Lightweight Region Growing Boundary-Aware UNet for Skin Lesion Segmentation

Accurate melanoma segmentation in dermoscopic images is critical for computer-aided diagnosis, but remains challenging when models must operate under strict parameter and computational budgets, as imposed by edge and mobile medical devices. Under such constraints, explicitly modeling fine-grained lesion boundaries is difficult, as most boundary-aware or attention-based mechanisms introduce non-negligible computational overhead. To resolve this accuracy–efficiency conflict, we propose a boundary-aware lightweight segmentation framework embedding explicit contour priors into efficient convolutional operators. Specifically, a region growing–based boundary generator produces continuous and reliable boundary priors, enabling explicit contour guidance without additional learnable parameters. Based on these priors, we design boundary-guided spatial–channel convolution to suppress redundant features in early layers, and a group-shared boundary-aware translation-variant convolution to refine irregular contours in deeper semantic layers with minimal parameter growth. In addition, a compact Dynamic Tanh-based activation module adaptively rescales feature responses during training, improving robustness to low-contrast lesions with negligible overhead. Experiments on ISIC2017 and ISIC2018 show the proposed method achieves comparable or improved segmentation accuracy while reducing parameters and FLOPs substantially, highlighting a favorable trade-off for resource-constrained clinical deployment.