Artificial Intelligence for Melanoma Diagnosis: A Qualitative Synthesis

Abstract

Aims: To identify what the scientific literature presents regarding the accuracy of Artificial Intelligence (AI) in diagnosing melanoma, to compare it with the accuracy of dermatologists, and to identify possible limitations. Study Design: Review Place and Duration of Study: Brazil, 01/19/2024. Methodology: The methodology used was a literature review, in which the studies included in the analysis were collected from the PubMed, Scielo, Cochrane, and Science Direct databases. Inclusion and exclusion criteria were followed, after which the selected studies were classified by two investigators and analyzed through full-text reading. Results: Eleven studies were included in the review. Most of the target population analyzed in the studies was male (36.4%), the predominant age group was 60 years or older (36.4%), and the predominant race was White. Three main diagnostic methods were identified for training the AIs: dermatoscopy, spectroscopy, and total body mapping. Fourteen different AI algorithms were evaluated by the studies. The highest sensitivity achieved by dermatologists in the studies was 96.5%, while the algorithms reached 97%. On the other hand, in terms of specificity, dermatologists achieved a maximum value of 99%, higher than the 97.1% achieved by the Ais. Conclusion: AI demonstrated sensitivity values similar to those of dermatologists, surpassing them in overall values. On the other hand, the maximum specificity achieved by dermatologists was higher. Studies revealed biases in the databases, a lack of demographic information about the patients, dataset restrictions and a small number of participating specialists. Therefore, AI emerges as a potential complementary method for melanoma diagnosis rather than a replacement for specialists; however, more robust studies are needed for its implementation in clinical practice.