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

Area of Science:

• Medical Aesthetics
• Artificial Intelligence in Dermatology
• Computer Vision for Clinical Analysis

Background:

• Nasolabial fold (NLF) severity is a crucial marker of facial aging and a common focus for aesthetic interventions.
• The Wrinkle Severity Rating Scale (WSRS) is the standard clinical tool for grading NLF severity but suffers from subjectivity and inter-observer variability.

Purpose of the Study:

• To develop and validate DeepFold, a deep learning ensemble model for automated, objective, and interpretable grading of NLF severity based on the WSRS.
• To establish a reliable AI-driven method for assessing NLF severity, overcoming the limitations of current subjective scales.

Main Methods:

• A dataset of 6,718 facial images was curated and annotated using the WSRS by three plastic surgeons.
• A ResNet-50 architecture was employed, with an ensemble strategy using majority voting across three independently trained networks.
• Model training utilized focal loss for class imbalance and early stopping, with performance evaluated by accuracy, F1-score, and confusion matrix analysis.

Main Results:

• The DeepFold ensemble model achieved a validation accuracy of 0.917 and an F1-score of 0.917.
• DeepFold outperformed individual baseline models, including ResNet-50 (accuracy: 0.904) and SeResNet-50 (accuracy: 0.882).
• Ensemble methods demonstrated reduced prediction variance and improved robustness, particularly under conditions of class imbalance.

Conclusions:

• DeepFold offers a dependable and standardized method for assessing NLF severity.
• The model holds significant potential clinical value for aesthetic evaluations, treatment planning, and monitoring treatment outcomes.
• Automated NLF grading using DeepFold can enhance objectivity and consistency in clinical practice.