Training AI Models for Aesthetic Facial Evaluation: Focused Review and Framework to Mitigate Homogenizing Bias

Area of Science:

• Medical Artificial Intelligence
• Computer Vision
• Biomedical Ethics

Background:

• Artificial intelligence (AI) models are increasingly used in facial aesthetic surgery for attractiveness prediction and outcome simulation.
• Current AI models exhibit bias, inaccurately evaluating underrepresented populations and risking aesthetic homogenization.
• This conflicts with patient goals of preserving ethnic features.

Purpose of the Study:

• To examine bias across AI development stages in aesthetic facial evaluation.
• To propose a comprehensive framework for mitigating bias in AI for facial aesthetics.
• To provide guidance for developing fairer AI tools in aesthetic surgery and facial analysis.

Main Methods:

• Literature review of bias in AI development for facial aesthetics.
• Analysis of benchmark datasets (e.g., SCUT-FBP, Chicago Face Database) for demographic representation.
• Proposal of a 6-pillar framework integrating diverse data collection, fairness-aware training, and continuous monitoring.

Main Results:

• Existing AI datasets and training methods underrepresent diverse populations (older adults, non-White individuals).
• Current evaluation metrics often lack demographic stratification, masking biases.
• Individual mitigation strategies are insufficient; a holistic framework is needed.

Conclusions:

• A 6-pillar framework is proposed, covering data diversity, fairness-aware training, intersectional metrics, explainable AI, stakeholder engagement, and monitoring.
• This framework aims to guide AI developers and clinicians in creating equitable AI tools.
• The principles are applicable to broader facial analysis applications beyond aesthetic surgery.