Comprehensive Review on U-Net Architectures for Skin Lesion Segmentation and Its Variants

Abstract

Skin lesion, ranging from benign growths to malignant tumors, like melanoma, pose significant diagnostic challenges because of their different morphological features. Early and accurate segmentation of these lesions from medical images is critical for effective diagnosis and treatment. Traditional image processing approaches, such as, thresholding and edge recognition, often fail to capture the complexity and variability of skin lesions. In contrast, deep learning methods, particularly Convolutional Neural Networks (CNNs), have modernized this field by providing robust solutions. The encoder-decoder architecture of the U-Net, for example design and skip connections, has showed significant consequence in describing lesion limits accurately. This review systematically examines the current state of skin lesions segmentation using u-net and its variants. It also highlights the performance of different models using metrics such as Dice coefficient and Jaccard index, addressing challenges such as the need for extensive annotated datasets. Our findings suggest that deep learning models like U-net significantly enhance the segmentation of skin lesions, improving diagnostic accuracy and clinical consequences.