A Non-Invasive Approach for Facial Action Unit Extraction and Its Application in Pain Detection.

Area of Science:

• Medical research
• Computer vision
• Biometrics

Background:

• Patient data privacy is a major hurdle in medical research.
• Sharing facial images poses significant privacy risks due to AI misuse.
• Facial expressions offer valuable health and emotional insights but require privacy-preserving methods.

Purpose of the Study:

• To develop a privacy-preserving method for analyzing facial expressions using 3D facial landmarks.
• To detect action units (AUs) and their intensities without using identifiable facial images.
• To demonstrate the utility of AU detection in a downstream task like pain detection.

Main Methods:

• Extracted 3D facial landmarks from video recordings.
• Employed a lightweight neural network (NN) for AU detection and intensity estimation.
• Utilized detected AUs and intensities to train deep learning (DL) models for pain detection.

Main Results:

• Achieved a 79.25% F1-score for main AU detection.
• Estimated AU intensities with a Root Mean Square Error (RMSE) of 0.66.
• Reached 91.16% accuracy in pain detection, comparable to image-based methods.

Conclusions:

• 3D facial landmarks can be shared instead of images, maintaining high accuracy in AU detection.
• The proposed method offers a computationally efficient and privacy-preserving alternative for facial expression analysis.
• This approach facilitates valuable insights from facial expressions in medical research without compromising patient privacy.