Automatic identification of the temporal retinal nerve fiber raphe from macular cube data

We evaluated several approaches for automatic location of the temporal nerve fiber raphe from standard macular cubes acquired on a Heidelberg Spectralis OCT. Macular cubes with B-scan separation of 96-122 µm were acquired from 15 healthy participants, and "high density" cubes with scan separation of 11 µm were acquired from the same eyes. These latter scans were assigned to experienced graders for subjective location of the raphe, providing the ground truth by which to compare methods operating on the lower density data. A variety of OCT scan parameters and image processing strategies were trialed. Vertically oriented scans, purposeful misalignment of the pupil to avoid reflective artifacts, and the use of intensity as opposed to thickness of the nerve fiber layer were all critical to minimize error. The best performing approach "cFan" involved projection of a fan of lines from each of several locations across the foveal pit; in each fan the line of least average intensity was identified. The centroid of the crossing points of these lines provided the raphe orientation with an average error of 1.5° (max = 4.1°) relative to the human graders. The disc-fovea-raphe angle was 172.4 ± 2.3° (range = 168.5-176.2°), which agrees well with other published estimates.