A Systematic Review of Computer Vision and Deep Learning Techniques for Skin Lesion Analysis

Abstract

Skin cancer, being one of the most common and fastest growing forms of cancer across the globe, early diagnosis is critical in increasing patient mortality. Conventional diagnostic techniques depend upon dermatologist's skill, dermoscopic analysis and biopsy procedures which may take substantial time and resources. Automated skin lesion analysis and classification systems have been revolutionized by recent advances in artificial intelligence, computer vision, and deep learning. This review paper discusses up to date computer vision and deep learning approaches for skin lesion detection, segmentation, classification, and diagnosis. The survey focus on image preprocessing methods, feature extraction techniques, convolutional neural networks, hybrid deep learning models, transfer learning approaches and ensemble architectures used for dermatology imaging. Moreover, the review covers difficulties to be considered while implementing methods mentioned in this paper as well as future research directions and evaluation metrics. The results show that deep learning-based systems perform exceptionally well, in terms of both accuracy and efficiency at the identification of malignant skin lesions and thus can assist clinical decision-making to improve diagnostic reliability. Yet problems like dataset imbalance, model interpretability, computational complexity and clinical integration continue to plague researchers in the field. The intent of this review is to provide a holistic view on emerging concepts and techniques in the context of intelligent skin cancer diagnosis systems.