Optimizing Skin Lesion Segmentation with UNet and Attention-Guidance Utilizing Test Time Augmentation

To properly diagnose and screen for skin disorders, precise segmentation of the affected area is required. Segmenting skin lesions is difficult because of factors such as the form of the lesion, its proximity to surrounding skin, and the presence of hair. To effectively tackle this matter, we present a pioneering technique that utilizes deep learning to achieve comprehensive automation in skin lesion segmentation. This approach incorporates advanced preprocessing methodologies to enhance the overall performance. The architecture Attention-UNet is employed for the purpose of selectively segmenting images of skin lesions. The high detection rate of this design allows for precise melanoma segmentation in real-time. Finally, in the post-processing phase, we applied test-time augmentation (TTA) to further enhance segmentation precision. Both the ISIC 2017 and ISIC 2018 skin lesion datasets were used for training and testing the suggested technique. The model's mDSC, mIoU, recall, and precision were all 91.93%, 88.15%, 92.72%, and 91.15%, respectively, when trained independently on the ISIC 2017 datasets. Separate training on the ISIC 2018 datasets achieved similar results: 92.17% mDSC, 90.26% mIoU, 92.34% recall, and 92.01% precision. After being trained on a dataset consisting ISIC data from both 2017 and 2018, the proposed method achieved mDSC scores of 92.30% percent, mIoU scores of 90.06% percent, recall rates of 93.91% percent, and precision rates of 90.74% percent. Our model demonstrates superior FPS scores compared to other models, indicating effective real-time performance. Due to its scalability and reliability, the suggested methodological framework may be utilized to create a completely automated computer-aided skin lesion diagnostic system.