Abstract
Fixational eye movements (FEMs), especially microsaccades (MS), are promising biomarkers of neurodegenerative disease. In vivo images of the photoreceptor mosaic acquired using an Adaptive Optics Scanning Laser Ophthalmoscope (AOSLO) are systematically distorted by eye motion. Most methods to extract FEMs from AOSLO data rely on comparison to a motion-free reference, giving eye-position as a function of time. MS are subsequently identified using adaptive velocity thresholds (Engbert & Kliegl, 2003). We use computer vision and machine learning (ML) for detection and characterisation of MS directly from raw AOSLO images. For training and validation, we use Emulated Retinal Image CApture (ERICA), an open-source tool to generate synthetic AOSLO datasets of retinal images and ground-truth velocity profiles (Young & Smithson, 2021). To classify regions of AOSLO images that contain a MS, images were divided into a grid of 32-by-32-pixel sub-images. Predictions from rows of sub-images aligned with the fast-scan of the AOSLO were combined, giving 1ms resolution. Model performance was high (F1 scores >0.92) across plausible MS displacement magnitudes and angles, with most errors close to the velocity threshold for classification. Direct velocity predictions were also derived from regression ML models. We show that ML models can be systematically adapted for generalisation to real in vivo images, allowing characterisation of MS at much finer spatial scales than video-based eye-trackers.