Machine Learning and Deep Learning for Automated Skin Disease Diagnosis: A Critical Narrative Review of Algorithms, Datasets, Performance Evaluation, and Future Research Directions

Abstract

Skin diseases, ranging from common inflammatory dermatoses to life-threatening malignancies such as melanoma, represent one of the highest-volume diagnostic burdens in global healthcare, and their visual nature has made dermatology a leading application domain for artificial intelligence. Machine learning (ML) and deep learning (DL) algorithms, particularly convolutional neural networks trained on dermoscopic and clinical photography, have been proposed as scalable decision-support tools capable of extending specialist-level diagnostic capability to primary care and underserved settings.
This critical systematic review aims to synthesize and appraise the current evidence base for ML/DL applications to automated skin disease diagnosis, focusing on algorithmic architectures, benchmark datasets, reported diagnostic performance and, crucially, performance equity across skin tones, and to assess the field's readiness for clinical translation.
A narrative-quantitative approach was used to synthesize 20 peer-reviewed published primary model-development studies, systematic reviews, and a comparator meta-synthesis from breast cancer imaging AI, identified using a PRISMA-informed methodology and published between 2021 and 2025 (see Figure 1 for the evidence identification and composition flow).
CNN- and hybrid-based architectures consistently achieve classification accuracy of 85–100% and AUC > 0.99 on curated benchmark datasets such as ISIC and HAM10000 (Table 1; Figure 2), and some models significantly outperform classical machine learning baselines. In an extensive study of physician-AI collaboration across 39 countries, however, clear differences in diagnostic performance by light and dark skin were observed, and systematic reviews all highlighted the fact that the datasets were not balanced, with some classes under-represented or absent, and that external validation of the algorithms was limited, with near-exclusive use of a few major light-skin dermoscopic archives (Table 2; Figure 3).