AA-ResUNet: an automatic skin lesion segmentation method based on fine-grained encoder-decoder architecture

Automated segmentation of skin lesions is a critical task in the computer-aided diagnosis of skin cancer. This article presents a novel deep learning model, Attention-Attention Residual U-Net (AA-ResUNet), designed to improve the precision and effectiveness of dermatoscopic image segmentation. The proposed model integrates a dual attention mechanism, spatial and channel-wise, within a residual U-Net framework to enhance feature extraction and suppress irrelevant information to enable fine-grained segmentation. The performance of the model is validated on the human against machine with 10000 training images (HAM10000), international skin imaging collaboration (ISIC) 2018, ISIC2016, and Pedro Hispano hospital (PH2) dermatological datasets. The experiment results show that AA-ResUNet obtains a test accuracy of 98.8%, Dice Similarity Coefficient of 93.6%, and mIoU of 95.7%. The dual attention mechanism dramatically improves segmentation accuracy and boundary accuracy, with better generalized performance than the state-of-the-art, confirmed by computer simulations.