No abstract

Eye motion is a major confound for magnetic resonance imaging (MRI) in neuroscience or ophthalmology.Currently, solutions toward eye stabilisation include participants fixating or administration of paralytics/anaesthetics. We developed a novel MRI protocol for acquiring 3-dimensional images while the eye freely moves. Eye motion serves as the basis for image reconstruction, rather than an impediment.We fully reconstruct videos of the moving eye and head. We quantitatively validate data quality with millimetre resolution in two ways for individual participants. First, eye position based on reconstructed images correlated with simultaneous eye-tracking. Second, the reconstructed images preserve anatomical properties; the eye's axial length measured from MRI images matched that obtained with ocular biometry.The technique operates on a standard clinical setup, without necessitating specialized hardware, facilitating wide deployment. In clinical practice, we anticipate that this may help reduce burden on both patients and infrastructure, by integrating multiple varieties of assessments into a single comprehensive session. More generally, our protocol is a harbinger for removing the necessity of fixation, thereby opening new opportunities for ethologically-valid, naturalistic paradigms, the inclusion of populations typically unable to stably fixate, and increased translational research such as in awake animals whose eye movements constitute an accessible behavioural readout.

Eye-movement trajectories are rich behavioral data, providing a window onto how the brain processes information. Analyses of these trajectories can be automated and benefit from machine learning algorithms. Among those, deep learning has recently proven very successful, setting new state-of-art results in many computer vision applications, including medical diagnosis systems. In this paper, we address the challenge of diagnosing and quantifying signs of visuospatial neglect from saccadic eye trajectories recorded in healthy controls and in brain-damaged patients with spatial neglect. We show how machine learning techniques, such as deep networks, can predict the patient's status with unprecedented accuracy, benchmarking the algorithm prediction with structural Magnetic Resonance Images (MRI) of the patients' brain lesions and their Diffusion Tensor Imaging (DTI) tracts. Preliminary evidence of correlation between MRI data and the algorithm scores suggest that a quantitative prediction of the patients' impairment based only onto the behavioral data of eye trajectories seem possible, therefore opening to new horizons in the field of non-invasive diagnostics.