Automatic Skin Lesion Segmentation Using an Enhanced U-Net Framework

Abstract

The given project offers an enhanced version of a U-Net-based framework for segmenting skin lesions with high precision because it is the best measure in the identification of melanoma at an early stage. ISIC2016, ISIC2017, and ISIC2018 datasets were used, including thousands of dermoscopic images with respective binary masks. The workflow ensemble with complex preprocessing (image resize, normalization, contrast enhancement, and severe data augmentation) was carried out to minimize the risk of the model being sensitive to changes in the appearance of the lesions. The training of the U-Net model involved a hybrid of Dice loss and binary cross-entropy loss, where the training was initiated to get a well-balanced performance between boundary and pixel-level accuracy. Morphological post-processing was used to clean up the small artifacts, fill the gaps, and smooth the mask borders to make the initial predictions more sophisticated. Specific measures of medical images, e.g., the Dice coefficient, Intersection over Union (IoU), precision, recall, and specificity, were employed to assess model performance. The data indicated very high segmentation accuracy in each of the datasets, which illustrated the robustness of the method in difficult cases that included low-contrast and irregularly shaped lesions. The approach is, thus, suitable for computer-aided diagnosis systems, which leads to more available and efficient skin cancer detection.