Automated early detection of androgenetic alopecia using deep learning on trichoscopic images from a Korean cohort: a retrospective model development and validation study.

Abstract

Purpose: This study developed and validated a deep learning model for the automated early detection of androgenetic alopecia (AGA) using trichoscopic images, and evaluated the model's diagnostic performance in a Korean clinical cohort.

Methods: We conducted a retrospective observational study using 318 trichoscopic scalp images labeled by board-certified dermatologists according to the Basic and Specific (BASP) system, collected at Ewha Womans University Medical Center between July 2018 and January 2024. The images were categorized as BASP 0 (no hair loss) or BASP 1-3 (early-stage hair loss). A ResNet-18 convolutional neural network, pretrained on ImageNet, was fine-tuned for binary classification. Internal validation was performed using stratified 5-fold cross-validation, and external validation was conducted through ensemble soft voting on a separate hold-out test set of 20 images. Model performance was measured by accuracy, precision, recall, F1-score, and area under the curve (AUC), with 95% confidence intervals (CIs) calculated for hold-out accuracy.

Results: Internal validation revealed robust model performance, with 4 out of 5 folds achieving an accuracy above 0.90 and an AUC above 0.93. In external validation on the hold-out test set, the ensemble model achieved an accuracy of 0.90 (95% CI, 0.77-1.03) and an AUC of 0.97, with perfect recall for early-stage hair loss. No missing data were present, and the model demonstrated stable convergence without requiring data augmentation.

Conclusion: This model demonstrated high accuracy and generalizability for detecting early-stage AGA from trichoscopic images, supporting its potential utility as a screening tool in clinical and teledermatology settings.