Pixel Tampering: Does Face Redaction Harm Medical AI Performance?

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Data Privacy

Background:

• Balancing data sharing and patient privacy is critical in medical imaging research.
• Face redaction tools anonymize head CT scans but may affect deep learning (DL) model performance.
• Existing DL models for medical imaging require robust anonymization techniques.

Purpose of the Study:

• To evaluate an open-source face redaction tool for head CT scans.
• To assess the impact of image redaction on DL model performance for age prediction.
• To compare models trained on redacted versus non-redacted data.

Main Methods:

• A Kaggle competition crowdsourced age prediction models using 2377 redacted head CT studies for training and 148 for testing.
• Top-performing models were evaluated on both redacted and non-redacted test sets.
• Performance was measured using mean absolute error (MAE); statistical significance was assessed using paired t-tests.

Main Results:

• The two best models achieved MAEs of 2.8 and 3.4 years on redacted test data.
• On non-redacted test data, MAEs increased to 3.2 and 3.8 years, with one model showing a significant performance drop (p=0.038).
• No significant difference was found between models on the redacted test set (p=0.610).

Conclusions:

• Models trained on redacted head CT data demonstrate minimal performance decline when applied to non-redacted images.
• The developed face redaction tool enables secure medical data sharing with limited impact on DL accuracy.
• This approach supports enhanced data sharing for medical imaging research while preserving patient privacy.