Automated Quantification of Shedding Hairs from Smartphone Photographs Using an AI Tool for Alopecia Monitoring.

INTRODUCTION

Objective shed-hair counts can support clinical assessment, longitudinal monitoring, and evaluation of treatment response. Manual counting is time-consuming, whereas visual estimates vary between observers. We developed an artificial intelligence (AI) model to count shedding hairs from ordinary smartphone photographs.

METHODS

Thirty volunteers acquired 5630 home smartphone photographs of dark shedding hairs on light backgrounds, and paired manual counts served as the reference standard. Images were assigned by participant into training/internal validation (24 volunteers, 4509 images) and independent testing (6 volunteers, 1121 images). A weakly supervised density-regression model was trained with image-level counts. A blinded reader study compared AI with visual estimates from 15 readers across 56 paired image sets. A non-randomized sequential refined wash test (RWT) comparison evaluated completion and usable quantitative data in historical manual and subsequent AI-assisted cohorts.

RESULTS

In independent testing, AI achieved a mean absolute error (MAE) of 5.4 hairs and a mean absolute percentage error (MAPE) of 7.0%. The mean model-minus-reference bias was -0.8 hairs (95% LoA, -17.0 to 15.4 hairs). In the blinded comparison, AI had lower error than all visual estimation methods (MAE, 6.99 hairs; MAPE, 6.15%; P < 0.001). Median total AI workflow time was 24.5 seconds versus 123.0 seconds for manual counting. Usable RWT data were available for 14 of 48 patients (29.2%) in the historical manual cohort and 23 of 30 (76.7%) in the subsequent AI-assisted cohort (P < 0.001).

CONCLUSION

Under the evaluated dark-hair and light-background conditions, AI provided low-error counts and lower error than human visual estimation. The subsequent AI-assisted RWT cohort more often provided usable quantitative data than the historical manual cohort in this non-randomized sequential comparison. Prospective studies are needed to evaluate workflow feasibility and potential benefits for clinical monitoring and patient outcomes.