Deep Learning-Based Acne Severity Classification Using Standardized Facial Images of Japanese Patients.

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Acne vulgaris severity assessment is vital for effective treatment.
• Existing deep learning models often lack specificity for diverse ethnic skin types, like Japanese skin, which is prone to post-inflammatory hyperpigmentation (PIH).

Purpose of the Study:

• To develop and validate a deep learning model for automated acne severity classification in Japanese patients.
• To provide an objective tool supporting diagnosis and treatment decisions for acne vulgaris.

Main Methods:

• Trained an EfficientNet-B2 model on 349 facial images of Japanese acne patients.
• Utilized a two-stage curriculum learning strategy, class weighting, and data augmentation.
• Evaluated performance on a hold-out test set using Investigator's Global Assessment (IGA) scores (0-4).

Main Results:

• Achieved 90.0% overall accuracy and a 0.885 macro-average F1-score.
• Demonstrated perfect recall (1.000) for severe acne classes (IGA-3 and IGA-4).
• Obtained a macro-average receiver operating characteristic-area under the curve (ROC-AUC) of 0.998.

Conclusions:

• A deep learning model trained on a Japanese dataset accurately classifies acne severity.
• This tool can standardize clinical evaluation and support timely treatment, potentially mitigating PIH and scarring risks.