Current state of machine learning for non-melanoma skin cancer

Authors:
Publisher website
Google Scholar
Abstract
Background Machine learning (ML) has been increasingly utilized for skin cancer screening, primarily of melanomas but also of non-melanoma skin cancers (NMSC). Objective This study presents the first quantitative review of the success of these techniques in NMSC screening. Methods A primary literature search was conducted using PubMed, MEDLINE, and arXiv, capturing all articles involving ML techniques and NMSC screening. Results 52 articles were included for quantitative analysis, resulting in a mean sensitivity of 89.2% (n = 52, 95% confidence interval (CI) 87.0–91.3) and a mean specificity of 81.1% (n = 44, 95% CI 74.5–87.8) for ML algorithms in the diagnosis of NMSC. Studies were further grouped by skin cancer type, algorithm type, diagnostic gold standard, data set source, and data set size. Conclusion There is insufficient evidence to conclude that an ML algorithm is superior at NMSC screening than a trained dermatologist utilizing dermoscopy for either BCC or SCC. Given that the studies included in this review were performed in silico, further study in the form of randomized clinical trials are needed to further elucidate the role of NMSC screening algorithms in dermatology.

This publication has 4 references indexed in Scilit: