A Deep Learning-Based Ensemble Model for Automated Nasolabial-Fold Severity Grading

Nasolabial-fold (NLF) severity is a key indicator of facial aging and a frequent target in aesthetic treatments. The Wrinkle Severity Rating Scale (WSRS) is widely used for clinical grading but remains inherently subjective and vulnerable to inter-observer variability.

The authors of this study aim to develop and validate DeepFold, a deep learning-based ensemble model for automated, objective, and clinically interpretable grading of NLF severity based on the WSRS.

A dataset of 6718 facial images was constructed, including 1718 images from clinical outpatients and 5000 from the CelebA dataset. All images were split into left and right halves and annotated independently by 3 senior plastic surgeons using the WSRS. ResNet-50 served as the base model architecture, and an ensemble strategy was applied using majority voting over 3 independently trained networks. Model training used focal loss to address class imbalance and was conducted in PyTorch with early stopping based on validation loss. Performance was assessed using accuracy, F1 score, and confusion matrix analysis.

The DeepFold ensemble model achieved a validation accuracy and F1 score of 0.917, outperforming individual baseline models such as ResNet-50 (accuracy: 0.904) and SeResNet-50 (accuracy: 0.882). Ensemble strategies reduced prediction variance and enhanced model robustness under class imbalance.

DeepFold provides a reliable and standardized approach to NLF severity assessment, offering potential clinical value in aesthetic evaluation, treatment planning, and outcome monitoring.