Advanced Deep Learning Strategies for Skin Lesion Segmentation using R2U-Net

Skin cancer poses serious health hazards in worldwide especially melanoma, with its early detection being crucial for improving survival rates. This paper focuses on the application of advanced deep learning (DL) models for the segmentation of skin lesions. Among the models explored, R2U-Net is enhanced through the integration of recurrent and residual connections, demonstrating superior performance in accurately segmenting complex skin lesions. Other models evaluated include U-Net, Pyramid-UNet, FCN, and SegNet. The study utilizes the Melanoma Research Alliance Multimodal Image Dataset (MRA-MIDAS), pre-processed using median filtering, and split into an 80-20% training-testing ratio. Evaluation metrics such as the Jaccard Index, Dice Coefficient, Cohen's Kappa, and Rand Index were employed to assess model performance. The R2U-Net achieved a IoU of 98.92 and DSC of 99.45, significantly outperforming other models. The findings support the use of R2U-Net for high-precision medical image segmentation tasks, particularly in detecting melanoma and other skin cancers, which will help the physicians to take quicker decisions, in turn will improve the health conditions of the patients.