Enhancing facial feature de-identification in multiframe brain images: A generative adversarial network approach

Area of Science:

• Neuroimaging
• Computer Vision
• Medical Informatics

Background:

• Public brain imaging datasets are growing, necessitating advanced de-identification methods for privacy compliance.
• Existing de-identification techniques may not fully address the nuances of facial feature privacy in head scans.
• The need for privacy-preserving techniques is critical in brain science research involving sensitive patient data.

Purpose of the Study:

• To develop and evaluate a novel deep learning-based approach for de-identifying facial features in brain images.
• To enhance privacy safeguards for public brain imaging datasets using generative adversarial networks (GANs) and 3D U-Net.
• To specifically target and synthesize facial features like ears, nose, mouth, and eyes for de-identification.

Main Methods:

• Utilized a generative adversarial network (GAN) to synthesize novel facial features and contours.
• Employed a 3D U-Net model for precise detection of facial features (ears, nose, mouth, eyes).
• Trained and tested the model on a combined dataset of 560 head CT and head MRI images, focusing on partial head regions.

Main Results:

• Achieved 100% accuracy for nose, mouth, and eye detection on the training dataset.
• Reported ear detection accuracy of 85.03% (training), 65.98% (testing), and 100% (validation).
• Structural Similarity Index (SSIM) analysis showed varying degrees of similarity between raw and generated facial features.

Conclusions:

• The proposed partial head image de-identification method is effective and suitable for real-world imaging scenarios.
• This deep learning approach advances de-identification technologies, strengthening privacy in brain imaging research.
• The methodology holds significant potential for future clinical applications and enhancing data security in neuroscience.