Purchase this article with an account.
Pengxiao Zang, Jie Wang, Tristan Hormel, Liang Liu, David Huang, Yali Jia; Automated Segmentation of Peripapillary Retinal Boundaries in OCT Combining Convolutional Neural Network and Graph Search. Invest. Ophthalmol. Vis. Sci. 2019;60(9):1503. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Within the peripapillary retina, measurements of retinal layer thickness from structural optical coherence tomography (OCT) and analysis of capillary plexuses from OCT angiography (OCTA) are useful for the early detection of glaucoma. The utility of such measurements depends critically on accurate retinal layer segmentation. We propose an automated algorithm for segmenting the peripapillary retinal boundaries that combines fully convolutional neural network (FCN) and graph search.
One eye of 35 glaucoma and 25 healthy participants was scanned by a spectral-domain OCT/OCTA system using a 4.5 × 4.5 mm2 scan pattern. Each scan contains 304 B-frames, with 304 A-lines on each frame. The disc boundary and boundaries of 7 peripapillary retinal layers in each B-frame were manually segmented by a certified human grader. A U-Net FCN architecture was modified and trained based on B-frames from 7 glaucomatous and 7 normal eyes. Probability maps including the 6 main peripapillary retinal layers were generated from the rest of scans by the trained U-Net. The disc boundary of each scan was obtained based on corresponding probability maps. Following this determination, 180 radial B-frames including A-lines from the center of optic disc, along with corresponding probability maps, were generated by resampling the raster coordinate (Fig. 1). The graph search was performed on the radial B-frames to segment the retinal layers based on corresponding probability maps. Final boundaries were transformed to raster coordinates and compared with manual grading.
The Dice similarity coefficient (DSC) was calculated between the disc boundaries of our method and manual delineation. The DSC was 0.93 ± 0.04 in glaucomatous and 0.92 ± 0.07 in normal eyes. The absolute errors of peripapillary retinal layer boundaries, as measured by pixel difference between automated and manual segmentations was 3.63 ± 1.16 µm in glaucomatous and 2.94 ± 0.75 µm in normal eyes (Fig. 2). The difference values (automated minus manual) of nerve fiber layer (NFL) thickness between our method and manual delineation was -3.15 ± 4.22 µm in glaucomatous and -2.10 ± 2.36 µm in normal eyes.
There was excellent agreement between automated and manual segmentations. The resulting NFL thickness and capillary density measurements may be useful in glaucoma evaluation.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
This PDF is available to Subscribers Only