Related Experiment Video
Updated: Sep 14, 2025
10:17
Guidelines and Experience Using Imaging Biomarker Explorer IBEX for Radiomics
Published on: January 8, 2018
13.3K
Re-identification of patients from imaging features extracted by foundation models
Giacomo Nebbia1, Sourav Kumar1, Stephen Michael McNamara1
1Ophthalmology Department, University of Colorado Anschutz Medical Campus, Aurora, CO, USA.
NPJ Digital Medicine
|July 22, 2025
Summary
Foundation models for medical imaging pose re-identification risks. Imaging features can reveal patient identity and demographic data, impacting privacy in ophthalmology and radiology.
Area of Science:
- Artificial Intelligence
- Medical Imaging
- Data Privacy
Background:
- Foundation models are increasingly used in medical imaging.
- Potential privacy risks, such as patient re-identification, are not well understood.
Purpose of the Study:
- To assess if imaging features from foundation models enable patient re-identification.
- To determine if re-identification correlates with demographic feature prediction.
Main Methods:
- Utilized Colour Fundus Photos (CFP), Optical Coherence Tomography (OCT) b-scans, and chest X-rays.
- Trained deep learning models to evaluate re-identification and demographic prediction performance.
Main Results:
- Reported patient re-identification rates of 40.3% (CFP), 46.3% (OCT), and 25.9% (X-rays).
- Demonstrated varying demographic prediction performance based on re-identification status.
- Achieved high image-level re-identification performance (82.3% CFP, 93.9% OCT, 63.7% X-ray).
Conclusions:
- Imaging features from medical foundation models contain information enabling patient re-identification.
- These findings highlight significant privacy concerns in AI-driven medical imaging.
