Deep learning-based defacing tool for CT angiography: CTA-DEFACE.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Data Privacy

Background:

• Computed tomography angiography (CTA) analysis increasingly utilizes artificial neural network (ANN) tools.
• Growing reliance on ANN for CTA data necessitates enhanced patient data protection measures.
• Automated defacing of sensitive facial information in CTA datasets is crucial for privacy.

Purpose of the Study:

• To develop and validate an automated defacing pipeline for computed tomography angiography (CTA) data using an artificial neural network (ANN).
• To ensure robust de-identification of patient data in CTA scans while maintaining image integrity for analysis.
• To compare the performance of the developed ANN defacing model against existing public algorithms.

Main Methods:

• Retrospective analysis of multi-institutional CTA datasets (n=100) for training an ANN model.
• Annotation of facemasks and subsequent ANN model training and external validation (n=50).
• Utilized MTCNN for face detection and FaceNet for verification to assess image similarity post-de-identification, calculating Dice Similarity Coefficient (DSC).

Main Results:

• The CTA-DEFACE model achieved a high Dice Similarity Coefficient (DSC) of 0.94 ± 0.02 for soft face tissue segmentation on the test set.
• Benchmarking against a public algorithm revealed significantly lower face detection probability (p < 0.001) and reduced similarity to original CTA images (p < 0.001) with CTA-DEFACE.
• The model demonstrated robust and precise defacing capabilities, validated externally and publicly accessible.

Conclusions:

• The developed ANN model, CTA-DEFACE, provides an effective and automated solution for de-identifying CTA data.
• CTA-DEFACE significantly outperforms a publicly available defacing algorithm in terms of privacy preservation.
• The model's external validation and public accessibility support its reliable application in clinical and research settings.