Accurate segmentation of dermoscopic images is essential for early melanoma diagnosis, yet current methods remain limited. CNN-based models capture local details but lack global context, whereas Transformer-based approaches model long-range dependencies but often lose fine structures; both struggle with blurred boundaries and scale variations. To address these challenges, we propose IFGNet, a hybrid segmentation framework that integrates CNN-Transformer synergy, multi-scale convolution, and boundary-aware decoding. Specifically, our design combines local-global feature fusion, large-kernel parallel convolutions without dilation, and a boundary refinement strategy to enhance lesion consistency. Extensive experiments on the ISIC 2016, ISIC 2017, and ISIC 2018 benchmarks demonstrate that IFGNet consistently surpasses state-of-the-art methods in Dice and IoU. These results highlight the effectiveness of IFGNet for accurate high-resolution skin lesion segmentation, and suggest its potential for clinical computer-aided melanoma diagnosis.
Li et al. (Fri,) studied this question.