Toward a Universal Measure of Facial Difference Using Two Novel Machine Learning Models

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Currently, no universally accepted, objective method exists for measuring facial deformity.
• Assessing facial abnormalities often relies on subjective human evaluation.
• Developing quantitative methods is crucial for consistent diagnosis and treatment monitoring.

Purpose of the Study:

• To develop and validate two distinct machine learning (ML) methods for objectively measuring facial deformity.
• To generate numerical scores reflecting the severity of various facial conditions.
• To compare ML-derived scores with human assessments.

Main Methods:

• Object detection using Haar features: A model trained on normal faces identifies abnormalities via confidence scores.
• Deep learning with convolutional autoencoders: Reconstruction error from normal face regeneration indicates deformity.
• Validation involved comparing ML scores against human ratings from 80 subjects evaluating 60 facial deformity images.

Main Results:

• Both ML methods produced scores highly correlated with average human ratings (Pearson's correlation coefficient > 0.96, P < 0.00001).
• The developed techniques demonstrated computational efficiency, delivering results within 3 seconds.
• The ML scores proved to be a reliable gauge of facial abnormality.

Conclusions:

• The proposed ML models show significant potential for integration into accessible handheld clinical tools.
• Further development could enable multicenter studies for comparing outcomes across different facial conditions, treatments, and institutions.
• These AI-driven approaches promise to standardize the objective measurement of facial deformities.