Markerless three-dimensional analysis of mouse eye movements using stereo vision

Quantifying three-dimensional mouse eye rotations, especially torsion, often requires invasive implanted coils or ocular markers. We developed and validated a markerless stereo vision-based method to measure yaw, pitch, and roll eye rotations.

Two aligned cameras were geometrically calibrated. A convolutional neural network detected pupil edges, which were fitted with ellipses. Matching edge points were reconstructed in three dimensions to estimate the rotation center and compute yaw and pitch; roll was derived from torsional shifts in pupil-edge irregularities. Performance was assessed in a computer-generated three-dimensional eyeball model, in five C57BL/6 J mice during angular vestibulo-ocular reflex testing, and in ten C57BL/6 J mice during otolith-ocular reflex testing.

In the model experiments, estimated rotations closely matched imposed trajectories; when the eyeball model was rotated by 1° per frame, mean absolute errors were 0.05-0.12°/frame across the three axes. In vivo, horizontal angular vestibulo-ocular reflex gains increased with stimulus frequency, ranging from 0.47 to 0.95 at 20°/s and from 0.36 to 1.08 at 50°/s. Static otolith-ocular reflex gains ranged from 0.67 to 0.78 with low variability and consistent rotation axes.

Compared with implanted-coil or marker-based approaches, our method reduces ocular invasiveness and we verified that it provides reproducible three-dimensional eye-rotation measurements.

Markerless stereo vision provides precise, practical quantification of mouse three-dimensional eye movements for vestibular studies.