Abstract
Purpose :
To develop AI supported fully automatic measurement of the waist of the nerve fiber layer at the optic nerve head, angularly resolved from 3D OCT volumes of the optic nerve head.
Methods :
3D representations of the ONH were captured with swept source OCT (OCT-Triton, Topcon, Japan) from both eyes in 8 male and 8 female young adults without known ocular pathology. Altogether, 3 volumes were captured for both eyes at one occasion. The waist of the nerve fiber layer at the optic nerve head was measured as the Pigment epithelium central limit Inner limit of the retina Minimal Distance (PIMD). Data from 5 males and 5 females were used to train a U-net AI model against the truth defined by semi-automatic annotation. This rendered, 30 000 annotated locations to guide the model during training.
The coordinates for the surface of the inner limit of the retina were identified with AI. The OCT volume was then split into 500 radii centered on the center of the ONH. In each radius, the coordinates for the Optic nerve head Pigment Epithelium Central limit (OPCL) was located with AI. Thereafter, the coordinates for the Inner surface of the Retina Closest Point (IRCP) to OPCL were automatically detected. Finally, the distance between IRCP and OPCL as a function of angle in the frontal plane was calculated. The average rotation of PIMD(Angle) for the 3 volumes within eye was estimated by cross correlation. Each volume within eye was then phase corrected to average rotation of volumes within eye and finally PIMD was averaged at each of the 500 angles for the 3 volumes.
Results :
The trained model fully automatically identified PIMD(Angle) in all OCT volumes that were not used for training the model. Cartesian and polar plots of PIMD(angle) in each eye revealed the expected double hump pattern with a maximum close to superiorly and inferiorly, respectively. A 95 % confidence interval for PIMD-2pi was estimated to 384 +/-68 um (d.f = 5).
Conclusions :
AI supported fully automatic estimation of PIMD(Angle) within an eye opens for early morphometric detection of glaucoma in a specific frontal plane angular segment. The wide limits of the confidence interval for PIMD-2pi reflects substantial variation in average nerve fiber layer thickness among subjects and indicate that it is more efficient to compare an eye to itself over time than to compare instant measurements to a normative database for subjects.
This is a 2020 ARVO Annual Meeting abstract.