Abstract
Current detection methods of Demodex mite density in facial erythema are semi-invasive or operator-dependent. We developed and evaluated a deep learning model (DemodexNet) for predicting Demodex mite density and assessed its impact on the diagnostic performance of dermatologists. This study included 1,124 patients with facial erythema who underwent Demodex mite density measurement at two referral hospitals between January 2016 and August 2023. DemodexNet achieved area under the receiver operating characteristic curve values of 0.823-0.865 in internal testing, with lower values observed in the external testing set. AI-assisted evaluation was associated with an increase in diagnostic accuracy among dermatologists from 63.7% to 70.6% (P < .001). Less experienced dermatologists and those with higher trust in AI showed greater performance gains. The model recognized central facial regions and individual lesions characteristic of demodicosis. DemodexNet demonstrates promising performance in predicting Demodex mite density and significantly improves dermatologists' diagnostic accuracy. As this proof-of-concept study was limited to Korean patients with Fitzpatrick skin types III-IV, validation in diverse populations is required before broader clinical application.