Optimizing MR-based gaze-decoding for eyes-closed eye-tracking in fMRI

Abstract

Eye movements provide valuable insights into human cognition and are a critical variable in numerous functional magnetic resonance imaging (fMRI) studies. Yet, when the eyes are closed, camera-based eye-tracking is unavailable, making studies of eyes-closed states challenging. Here, we address this gap using DeepMReye, a deep learning framework for camera-free gaze reconstruction from the MR-signal of the eyes. We first show that fine-tuning DeepMReye on visuomotor calibration data acquired with the eyes open significantly improves gaze decoding, and that this fine-tuning does not require simultaneous camera-based data. We next assessed whether decoding could be extended to eyes-closed states using a novel auditory-guided task, in which participants gazed at learned target positions with and without visual input, and with their eyes open, blinking, or closed. While DeepMReye was originally trained exclusively on eyes-open data, the network successfully generalized to eyes-closed periods, and this generalization was further improved through task-specific fine-tuning. Finally, fine-tuning also improved decoding of eyelid-state (open versus closed) directly from the MR-signal. These findings demonstrate that gaze and eyelid-state monitoring during eyes-closed periods is feasible, enabling broader integration of eye-tracking in fMRI research.