The rapid development of computer-aided diagnosis technology has significantly improved the efficiency of skin lesion screening. However, the automatic segmentation of skin lesion remains challenging due to various interfering factors in dermoscopic images. Models based on Skip-Connecting have demonstrated strong performance in segmenting small targets and fuzzy boundaries, and those utilizing Atrous-Convolution excel in handling large-scale and multi-scale lesion segmentation, establishing these as two predominant approaches in skin lesion image segmentation. Furthermore, the Transformer model, renowned for its powerful global modeling and long-distance dependency capture capabilities, is poised for deeper integration with the traditional deep learning model to further optimize the segmentation performance. Accordingly, this review systematically outlines the application progress of the Skip-Connecting model and the Atrous-Convolution model in skin lesion segmentation and the improvement strategy of their derivative network. It focuses on analyzing the application paradigm of the Transformer model in this field and its coupling method with traditional models. Exploratory research on emerging frameworks, such as the visual state space model (Mamba), the generative confrontation network, and the diffusion model, is also examined. Finally, in light of the limitations in clinical research and the underlying causes of suboptimal segmentation effect, corresponding solutions are proposed, and future research directions are prospected.