FDS-Net: Frequency-Domain-Driven Multi-Scale Supervised Contrastive Learning Network for Skin Cancer Segmentation

The low visual contrast between dermatological lesions and normal skin, coupled with the masking effect of surface artefacts, presents two major obstacles to precise automated skin lesion segmentation. In order to address these challenges, the present paper puts forward a frequency-domain-driven multi-scale supervised contrast segmentation network (FDS-Net). Firstly, a wavelet-based semantic information decoupling module was designed. This module separates image frequency-domain information into lesion, normal, and boundary regions, enabling efficient extraction of target frequency bands. Subsequently, a feature fusion enhancement module was introduced. This module has been shown to capture critical dual-domain information through a dual-branch architecture. This strengthens the discriminative power between lesion and normal regions. Consequently, a Mahalanobis distance-based progressive decoder was developed. The implementation of Mahalanobis distance similarity metrics has been demonstrated to enhance the model&#x2019;s capacity to manage blurred boundaries. Experiments on three public datasets demonstrate optimal performance across all tests. Specifically, on the ISIC2018 dataset, the proposed method attained 82.42% mIOU and 89.51% mDSC, thereby substantiating its superiority and universality. Our released code is available at <uri>https://github.com/zlw-wsy/FDS-Net</uri>