Multi-Scale Perceptual Learning for Skin Lesion Image Segmentation

Skin lesion segmentation is a crucial task in computer-aided diagnosis for early detection of skin diseases, including melanoma. However, accurately segmenting skin lesions in images remains a challenging problem due to variations in lesion size, shape, and texture. To address these challenges, we propose a novel multi-scale perceptual learning method for skin lesion image segmentation, referred to as Skin-MPL. Specifically, we first utilize the Mix Transformer as an encoder to extract multi-scale features across four stages. A Receptive Field Enhancement (RFE) module is then employed to improve the receptive field of these four feature scales. Next, these multi-scale features are fed into a Pyramid Decoder (PD) to generate an initial coarse map. Finally, we design a Multi-Scale Convolution Fusion (MCF) module along with a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$3 \times 3$</tex> convolution layer for feature fusion and decoding. Experimental results on the publicly available ISIC2017 and ISIC2018 datasets demonstrate that our method achieves superior performance, outperforming existing approaches in the task of skin lesion segmentation. The proposed method effectively captures multi-scale contextual information, offering significant improvements in segmentation accuracy.