An improved lightweight segmentation neural network for dermoscopic lesion images based on knowledge distillation

As artificial intelligence technology rapidly advances, numerous medical image segmentation models based on deep learning have been developed. However, the current mainstream neural network algorithms usually have too many parameters, complex models, and high computational complexity. To better facilitate the operation of neural networks on low-power devices while preserving model segmentation precision, this paper introduces an enhanced lightweight segmentation neural network for dermoscopic lesion images, LDLSNet (Lightweight Dermoscopic Lesion Segmentation Net), by redesigning the network structure and employing advanced attention mechanisms. The model’s parameter count is reduced via knowledge distillation. LDLSNet can address the limitations of current skin lesion images segmentation techniques and offers a more lightweight approach to segmentation. This paper elaborates on the architecture and learning methodology of the suggested LDLSNet model and carries out experimental substantiation using the ISIC dataset. Experimental results indicate that while maintaining superior performance in dermoscopic lesion images segmentation tasks, the proposed method possesses fewer network parameters. This research offers a novel and effective solution in the field of lightweight dermoscopic lesion images segmentation, holding certain theoretical and practical implications.