Abstract
Purpose :
Direct observation is typically used to assess the ocular range of motion, and while valuable, this subjective approach does not allow ocular rotations to be quantified. We developed a novel image processing technique, named limbus segment tracking (LST), to address this methodological gap.
Methods :
LST uses the location and shape of limbal segments to estimate the magnitude and direction of ocular rotations. This is done by “linking” rotation parameters to limbus boundaries extracted from ocular photographs using a combination of machine learning and mathematical optimization. Instantiation of LST’s components proceeded in two steps. First, we captured ocular photographs from head-stabilized observers gazing in different directions. The images were then annotated and used to train a SOLO v2.0 network to segment both the iris and sclera. Limbal boundaries, defined as the juxtaposition of the iris and sclera, were then extracted from a set of independent test images segmented by the trained network. Second, mathematical optimization was used to find the rotation parameters which best matched the location and shape of each extracted limbus segment. The precision and recall of the neural network, in conjunction with the optimization error, were used to evaluate segmentation and optimization performance, and thus, the potential clinical applicability of our technique.
Results :
The overall precision and recall of the trained SOLO v2.0 network were 95% and 93.61%, respectively. When examined individually, iris segmentation tended to produce higher values for both metrics (+1.78% and +2.57% vs. scleral segmentation). The mean optimization error, comparing the estimated to extracted limbus segments, was 0.15 mm. The precision, recall, and optimization error were all independent of the estimated magnitude and direction of rotation.
Conclusions :
LST’s constituent components transformed ocular photographs into estimated ocular rotations with ideal segmentation performance and low optimization error. The next steps are to assess this technique’s validity through comparison of rotation estimates to known target locations.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.