Abstract
BACKGROUND: Accurate severity assessment of acne vulgaris is crucial for treatment selection and monitoring. While deep learning models have shown promise, most are based on Caucasian datasets, and models specific to Japanese patients, who have distinct skin characteristics such as a higher propensity for post-inflammatory hyperpigmentation (PIH), are lacking. OBJECTIVE: To develop and internally validate a deep learning model for automated acne severity classification using facial images of Japanese patients labeled with the Investigator's Global Assessment (IGA) score, aiming to provide an objective tool to support diagnosis and treatment decisions. METHODS: A dataset of 349 facial images was collected from Japanese acne patients. After preprocessing, including region of interest (ROI) extraction, images were labeled with IGA scores (0-4) by board-certified dermatologists. An EfficientNet-B2 model was trained using a two-stage curriculum learning strategy, class weighting, and extensive data augmentation techniques. The model's performance was evaluated on a hold-out test set. RESULTS: The model achieved an overall accuracy of 90.0% and a macro-average F1-score of 0.885 on the test set. Notably, it demonstrated perfect recall (1.000) for severe acne classes (IGA-3 and IGA-4), indicating exceptional performance in identifying patients requiring prompt therapeutic intervention. The macro-average receiver operating characteristic-area under the curve (ROC-AUC) was 0.998. CONCLUSION: Our deep learning model, trained on a dedicated Japanese dataset, can classify acne severity with high accuracy. This tool has the potential to support objective clinical assessment, standardize evaluation, and potentially contribute to reducing the risk of PIH and scarring by facilitating timely and appropriate treatment.