Backgrounds: Establishment of quantitative, objective evaluation tool for facial palsy has been a big issue for clinicians and researchers, and in an era of artificial intelligence, AI-driven video analysis can be considered a reasonable solution for this long-discussed issue. We introduced facial keypoint detection, which detects facial landmarks with 68 points, but existing models had been organized almost solely with images of healthy individuals and low accuracy was presumed in the prediction of asymmetric faces of facial palsy patients.The accuracy of existing model was assessed through applying it to the movies of 30 facial palsy patients. Qualitative review clearly showed its insufficiency; prone to detect patients' faces as symmetric, and unable to detect closure of the eyes. Thus, we decided to enhance the model through machine learning process of annotation (fine-tuning).
Methods: 1181 images extracted from the movies of 196 patients were enrolled in the training, and these images underwent manual correction of 68 keypoints. The annotated data was integrated into the previous model with a stack of two Hourglass networks combined with Channel Aggregation Block.
Results: The post-annotation model showed improvement in normalized mean error from 0.026 to 0.018, and also qualitatively keypoint detection on each facial unit revealed improvements.
Conclusion: Our strict control of the inter- and intra-annotator variability successfully fine-tuned the presenting model, and we consider our new model as one of promising solutions for objective assessment of facial palsy.
Copyright © 2024 by the American Society of Plastic Surgeons.