fMRI-Based Prediction of Eye Gaze During Naturalistic Movie Viewing Reveals Eye-Movement-Related Brain Activity.

Eye gaze and eye movements provide important indices of perceptual and cognitive processes, particularly under naturalistic conditions such as movie viewing. However, concurrent eye tracking is often unavailable in functional MRI (fMRI) studies due to technical and logistical constraints. Recent deep learning approaches have made it possible to estimate eye gaze directly from eyeball signals in fMRI data, offering a potential alternative to camera-based eye tracking. Here, we applied a pre-trained fMRI-based deep neural network model (DeepMReye) to estimate eye gaze during movie watching across three independent fMRI datasets. Model performance was evaluated by comparison with camera-based eye-tracking data when available, as well as by assessing inter-individual correlations, a commonly used benchmark in naturalistic fMRI research. At the individual level, predicted gaze showed modest correspondence with measured data (r ≈ -0.38 to 0.67). In contrast, group-averaged gaze predictions exhibited substantially higher correlations (r ≈ 0.7-0.8), indicating improved reliability at the group level. We further derived eye-movement-related time series from the predicted gaze signals and examined their associated brain activity. Consistent with differences in prediction accuracy, individual-level analyses yielded activation patterns largely restricted to visual cortex, whereas group-averaged predictions revealed more widespread activation, including established oculomotor control regions such as frontal and parietal eye fields. Exploratory analyses indicated age-related effects on gaze prediction accuracy and eye-movement-related brain activity, although these effects were not consistent across datasets. Together, these findings demonstrate that group-averaged fMRI-based gaze estimation can support the investigation of eye-movement-related brain activity in naturalistic paradigms, while highlighting current limitations for individual-level inference. The results provide a methodological assessment of fMRI-based gaze prediction and inform its appropriate use in future neuroimaging studies.