Automated Segmentation of the Anterior Lamina Cribrosa Surface (ALCS) within Optic Nerve Head (ONH) Optical Coherence Tomography (OCT) B-scans

Abstract

Purpose : To characterize the difference between automated and manual segmentation of the ALCS within ONH OCT B-scans.

Methods : One eye from 397 normal and 342 glaucoma subjects and suspects underwent OCT enhanced depth imaging (Heidelberg Engineering Spectralis) using a 24 radial B-scan pattern. Manual ALCS segmentations were performed using custom software. Approximately 50% of the eyes (198 normal/196 glaucoma) were used to train an automated segmentation algorithm (Spectralis SP-X software 6.4.8.116). The remaining 50% were used to test its performance. Within each B-scan, ALCS z-axis vertical depth (ALCS-Z) relative to the top of image was computed for each pixel at which ALCS could be identified, averaged for each segmentation method, and the absolute difference between the methods (Δ ALCS-Z) was computed. In 16 eyes, ALCS was independently segmented by 4 operators. Bland-Altman limits of agreement among operators (LOA-O) and between the average of the 4 operators (operator average) and automated segmentation (LOA-OA) for ALCS-Z were calculated. The relationship between glaucoma status and Δ ALCS-Z was explored using a General Estimation Equation model.

Results : The final testing dataset included 8653 B-scans with visible ALCS from 361 subjects (191 normal, 170 glaucoma, median age 65, range 20-95). The median (interquartile range) Δ ALCS-Z value was 9.8 (4.6, 16.7) µm in glaucoma and 11.6 (5.6, 21.2) µm in normal eyes. Δ ALCS-Z was greater in normal compared to glaucoma eyes (p <0.0001), however the median differences in both groups were <3 image pixels. Within the 16 reproducibility eyes, the mean ALCS-Z difference between the automated and 4 operator average segmentation values was -1 µm. While the ALCS-Z LOA-OA were -27 to +26 µm, the automated segmentation value was within the LOA-O (-20, +20) in 96.6% of the B-scans (Fig 1). Within the 361 study eyes, 89.4% of automated segmentations fell within the LOA-O (-28 to +28, after adjusting to account for single operator delineation, Fig 2).

Conclusions : Automated ALCS segmentation within 361 normal, glaucoma suspect and subject eyes yielded ALCS z-axis depth values that fell within the between-operator limits of agreement (LOA-O) in 89.4% of B-scans. These findings support further development of automated ALCS segmentation for the detection of glaucoma and its progression.

This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.