A Lightweight Statistical Multi-Feature Adaptive Attention Network for Dermoscopic Image Segmentation

Abstract

With the advent of Transformer architectures, the segmentation performance of dermoscopic images has been significantly enhanced. However, the substantial computational load associated with Transformers limits their feasibility for deployment on resource-constrained mobile devices. To address this challenge, we propose a Statistical Multi-feature Adaptive Attention Network (SFANet) that aims to achieve a balance between segmentation accuracy and computational efficiency. In SFANet, we propose a Multi-dilation Asymmetric Convolution Block (MDACB) and a Group Feature Mask Enhancement Component (GMEC). MDACB is composed of Multi-dilation Asymmetric Convolution (MDAC), a set of ultra-lightweight Statistical Multi-feature Adaptive Spatial Recalibration Attention (SASA) modules, Statistical Multi-feature Adaptive Channel Recalibration Attention (SACA) modules, and residual connections. MDAC efficiently captures a wider range of contextual information while maintaining a lightweight structure. SASA and SACA integrate multi-statistical features along spatial and channel dimensions, adaptively fusing mean, maximum, standard deviation, and energy via learnable weights. Convolution operations then model spatial dependencies and capture cross-channel interactions to generate attention weights, enabling precise feature recalibration in both dimensions. GMEC groups features from lower decoding layers and skip connections, and then merges them with the corresponding stage-generated masks, enabling efficient and accurate feature processing in the decoding layers while maintaining a low parameter count. Experiments on the ISIC2017, ISIC2018, and PH2 datasets demonstrate that SFANet achieves a mIoU of 80.15%, 81.12%, and 85.30%, with only 0.037 M parameters and 0.234 GFLOPs. Our code is publicly available at https://github.com/cwy1024/SFANet.