Purchase this article with an account.
Jenwei Kuo, Qi Yang, Wei Chieh Huang, Charles A Reisman; Automated Bruch’s membrane segmentation in OCT volumes with choroidal neovascularization.. Invest. Ophthalmol. Vis. Sci. 2019;60(9):152.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To present an enhanced Bruch’s Membrane (BM) segmentation algorithm for optical coherence tomography (OCT) 3D volumes with choroidal neovascularization (CNV), thus improving optical coherence tomography angiography (OCT-A) slab visualizations.
The proposed segmentation algorithm extends the legacy segmentation algorithm used in DRI OCT Triton (Topcon Corp., Tokyo, Japan) to further refine the segmentation result of BM when a large detachment of the retinal pigment epithelium (RPE) is detected. The proposed algorithm consists of two steps. First, a series of image processing techniques including image pre-processing, gradient extraction, and edge detection were applied to obtain candidate edge segments in the target 3D volume. Second, BM edge segments were determined by feature analysis, and the final BM boundary was obtained by 3D fitting with both determined BM edge segments and the result from the legacy algorithm. The automated segmentation results were compared to manually delineated BM boundary obtained from two 3x3 mm (320 a-lines x 320 b-scans) and two 6x6 mm (512x512) macula OCT-A scans acquired by one-micron wavelength swept-source OCT (DRI OCT Triton). Signed and unsigned border position differences (mean and standard deviation) of the BM at 21 fovea center B-scans were calculated for segmentation accuracy evaluation. The OCT-A projections were also visually checked and compared.
Signed differences (in µm) between manual segmentation result and the proposed algorithm in 3x3 and 6x6 images are 6.15 ± 8.12 and -1.44 ± 7.75, respectively. Unsigned differences in 3x3 and 6x6 images are 7.69 ± 6.69 and 5.38 ± 5,76, respectively. Positive differences indicate the automated segmentation result is more vitreal (i.e. closer to RPE) to the manual segmentation. The improvement in CNV visualization in OCT-A is shown in Fig. 1.
The proposed automated BM detection routine improves CNV visualization in OCT-A. The proposed method can be incorporated into the legacy segmentation as an additional feature to generate reliable segmentation results for clinical interpretation.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
Figure 1. OCT-A en face images (integration from RPE to BM) for two fully automated BM detection routines. (a) legacy segmentation; and (b) enhanced methodology.
This PDF is available to Subscribers Only