Abstract
Skin scars, resulting from the natural healing cascade following cutaneous injury, impose enduring physiological and psychological burdens on patients. This review first summarizes the biological classification of scars, their formation mechanisms, and conventional clinical assessment techniques. We then introduce core concepts of artificial intelligence, contrasting traditional machine learning algorithms with modern deep learning architectures, and review publicly available dermatology datasets. Standardized quantitative evaluation metrics and benchmarking protocols are presented to enable fair comparisons across studies. In the Methods Review section, we employ a systematic literature search strategy. Traditional machine learning methods are classified into unsupervised and supervised approaches. We examine convolutional neural networks (CNNs) as an independent category. We also explore advanced algorithms, including multimodal fusion, attention mechanisms, and self-supervised and generative models. For each category, we outline the technical approach, emphasize performance benefits, and discuss inherent limitations. Throughout, we also highlight key challenges related to data scarcity, domain shifts, and privacy legislation, and propose recommendations to enhance robustness, generalizability, and clinical interpretability. By aligning current capabilities with unmet clinical needs, this review offers a coherent roadmap for future research and the translational deployment of intelligent scar diagnosis systems.