Publication
Revealing Eye-dentity: Foundation Models Enable Re-identification from Retinal OCT
Marc Seibel; Nele Brügge; Timo Kepp; Bennet Kahrs; Jan Ehrhardt; Heinz Handels
In: Heinz Handels; Katharina Breininger; Thomas Deserno; Andreas Maier; Klaus Maier-Hein; Christoph Palm; Thomas Tolxdorff (Hrsg.). Bildverarbeitung für die Medizin 2026 Proceedings, German Conference on Medical Image Computing. Workshop Bildverarbeitung für die Medizin (BVM-2026), Lübeck, Germany, Pages 63-69, ISBN 978-3-658-51100-5, Springer Vieweg, Wiesbaden, 3/2026.
Abstract
Foundation models have become central to medical imaging research, yet subject re-identification implications remain unclear. In this work, we study whether optical coherence tomography (OCT)-derived B-scan features extracted using frozen generalist and specialist foundation models allow re-identification of subjects intra- and cross-device. A lightweight binary classifier was trained to predict whether two feature sets originate from the same individual. Results show that specialist models such as RETFound reach 78% re-identification accuracy (Rank-1) on high-resolution OCT data, while generalist models perform only slightly worse. Performance decreased substantially on the lower-resolution data and was near chance across devices. These findings suggest that general foundation models extract subject-related information, potentially entangled with recording device-related information.