Title : Outcomes of artificial intelligence use for analysis of non-invasive dermatological images, comparison of skin datasets and performance versus humans
Abstract:
In the sector of non-invasive skin diagnostics, artificial intelligence trained using datasets consisting of multiple clinical gross images, dermoscopy images, etc. of normal and pathological skin conditions collected from different hospitals around the world by doctors have shown different degrees of accuracy and reliability in diagnosing skin conditions using the non-invasive dermatological images provided. This study aims to review the outcomes of AI use in the analysis of non-invasive dermatological images of different skin diseases, comparing studies with AI trained with different skin datasets and the performance of AI versus humans. Two databases, Pubmed and Cochrane Library database were used to identify studies for this review by devising a search strategy using appropriate keywords and PRISMA protocols. This resulted in a total of 16 studies to be included in our review. The results are discussed under three subtopics- diagnostic accuracy of AI analysis, comparison of different algorithms, and performance of AI versus humans. The limitations of the study along with areas of caution were also noted. In conclusion, this review shows that AI diagnosis can be accurate, more or equal to that of trained dermatologists and can vary according to algorithms, datasets used and augmentations applied to the data. It is also essential to continue to improve the standard of care and in some cases, can also improve human performance when AI data is used to further their diagnostic accuracy.
