Deep Learning for Diagnosis of Tinea Corporis and Tinea Cruris

Abstract

Background:

Tinea corporis and tinea cruris are frequently misdiagnosed, and studies utilizing deep learning for their diagnosis remain scarce.

Objectives:

This study primarily aimed to develop a deep learning model trained on images from diverse clinical settings to distinguish dermatophytosis from non-dermatophytosis. A secondary objective was to compare its diagnostic performance with that of physicians.

Methods:

This retrospective diagnostic study analyzed clinical images of dermatophytosis (tinea corporis/cruris) and non-dermatophytosis (eczema, psoriasis, and lichen planus). Dermatophytosis was confirmed by the presence of hyphae, whereas non-dermatophytosis was diagnosed based on clinical and histological findings. The deep learning model utilized a multitask learning approach, integrating classification and segmentation. Adaptive weighting facilitated area-of-interest segmentation, improving attention to relevant features. Model performance was compared with physicians using a 20-image quiz from the test dataset.

Results:

A total of 1400 images (600 dermatophytosis, 840 non-dermatophytosis) from 580 Thai patients, predominantly with Fitzpatrick skin types III and IV, were analyzed. In the test dataset, the model achieved an AUC of 0.84 (95% CI: 0.80–0.88), a sensitivity of 0.80 (95% CI: 0.74–0.85), a specificity of 0.71 (95% CI: 0.65–0.76), an accuracy of 0.74 (95% CI: 0.70–0.78), and an F1-score of 0.71 (95% CI: 0.66–0.76). AUC values for deep learning, dermatologists (n = 15), and non-dermatologists (n = 11) on the image quiz were 0.80, 0.78, and 0.75, respectively.

Conclusions:

The deep learning model demonstrated effective diagnostic performance for tinea corporis/cruris in Asian skin types, performing comparably to dermatologists as evaluated using a 20-image quiz derived from the test dataset.