Unsupervised anomaly appraisal of cleft faces using a StyleGAN2-based model adaptation technique

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Congenital cleft lip anomalies present a significant clinical challenge for objective assessment.
• Existing methods for measuring deformity lack clinical feasibility and objectivity.
• There is a need for reliable tools to quantify baseline deformity and surgical outcomes.

Purpose of the Study:

• To introduce a novel machine learning framework for detecting, localizing, and measuring cleft lip anomaly severity.
• To provide an objective and clinically feasible method for assessing facial deformity and surgical changes.
• To establish a new standard for automated, real-time clinical measurement of cleft lip.

Main Methods:

• Utilized StyleGAN2 generative adversarial network for face normalization.
• Employed a pixel-wise subtraction approach to quantify deformity severity.
• Developed a pipeline including image preprocessing, heat-map generation, and abnormality scoring.

Main Results:

• The framework demonstrated high correlation with human ratings (Pearson's r = 0.89).
• The pixel-wise measurement technique outperformed existing state-of-the-art image quality metrics.
• Heatmaps visually corroborated the generated anomaly scores, highlighting anatomical anomalies.

Conclusions:

• The proposed machine learning framework offers an objective and automated solution for measuring cleft lip severity.
• This technology has the potential to become a new standard in clinical assessment of congenital facial deformities.
• The model provides a feasible method for tracking changes related to reconstructive surgery.