SGT-Net: Spectral generalized transformer for skin lesion segmentation

Skin lesion segmentation is vital for melanoma diagnosis. Existing deep-learning models face issues like poor generalization due to image noise and unique-shaped images. This paper presents the Spectral Generalized Transformer, a new network based on the classic encoder–decoder. It includes a redesigned Spectral Adaptive Block (SAB) for better feature capture and connection. SAB uses Fourier Transform to convert features, extracts key features by frequency. A Feature-generalization Decoder (FD) is developed for spectral features. With the reverse attention mechanism and multi-scale fusion, it enhances edge segmentation and model generalization. Evaluated on four public datasets, the model’s mIou on PH2, ISIC2016, ISIC2017, and ISIC2018 reached 92.28%, 87.01%, 83.61%, and 81.52% respectively. The results show it surpasses state-of-the-art methods. Code available at: https://github.com/1914010125/Zeye-Xu.