ARCUNet: enhancing skin lesion segmentation with residual convolutions and attention mechanisms for improved accuracy and robustness

Skin lesion segmentation presents significant challenges due to the high variability in lesion size, shape, color, and texture and the presence of artifacts like hair, shadows, and reflections, which complicate accurate boundary delineation. To address these challenges, we proposed ARCUNet, a semantic segmentation model including residual convolutions and attention techniques to improve segmentation accuracy to address the challenges of skin lesion segmentation, By incorporating residual convolutions and attention mechanisms, ARCUNet enhances feature learning, stabilizes training, and sharpens focus on lesion boundaries for improved segmentation accuracy. Residual convolutions ensure better gradient flow and faster convergence, while attention mechanisms refine feature selection by emphasizing critical lesion regions and suppressing irrelevant details. The model was tested on the ISIC 2016, 2017, and 2018 datasets with outstanding segmentation results with accuracy measures of 98.12%, 96.45%, and 98.19%, Dice measures of 94.68%, 91.21%, and 95.34%, and Jaccard measures of 91.14%, 88.33%, and 93.53%, respectively. These findings signify the ability of ARCUNet to segment skin lesions accurately and thus as an effective tool for computerized skin disease diagnosis.