Analysis of trichoscopic images using deep neural networks for the diagnosis and activity assessment of alopecia areata - a retrospective study.

Abstract

Background and objectives: Alopecia areata (AA) is an autoimmune disease that provokes hair loss. The diagnosis is made clinically with the support of trichoscopy. However, trichoscopy requires specialized training. Deep learning models may support the diagnosis and management of AA. The aim of this study is to develop a deep learning framework to diagnose AA and to determine the AA level of activity.

Patients and methods: A retrospective analysis of trichoscopic images collected from patients with scalp diseases and healthy controls was conducted to develop a two-step deep learning framework. In Step-1, the model aimed to distinguish AA disease from both other scalp diseases and control healthy subjects. In Step-2, we intended to train a model that recognized the AA level of activity dividing the AA dataset into active, inactive, and regrowth.

Results: In Step-1 an overall accuracy of 88.92% and an F1 score of 88.17% were achieved with an AA discriminatory capacity of 90.98%. In Step-2 an accuracy of 83.33% and an F1 score of 83.36% were reached.

Conclusions: Our study highlighted for the first time the potential use of artificial intelligence in the diagnosis and staging of AA allowing more accurate diagnoses and better care.