Automated 3D Segmentation of Bruch's Membrane Opening from SD-OCT Volumes

Abstract

Purpose : Bruch's membrane opening-minimum rim width (BMO-MRW) is shown to be superior to other conventional structural parameters for diagnosing glaucoma. While most existing approaches either segment the 2D projection of BMO points or identify BMO points in individual slices, we propose a machine-learning graph based approach for 3D segmentation of BMO from glaucomatous spectral domain-optical coherence tomography (SD-OCT) volumes.

Methods : Forty-four optic-nerve-head-centered SD-OCT volumes were acquired from 44 human patients with varying stages of glaucoma using Cirrus SD-OCT (Carl Zeiss Meditec, Inc.). The BMO identification is formulated as an optimization problem for finding a 3D path within the SD-OCT volume. To this end, we transfer the volume to the radial domain such that the closed loop-path of BMO points in the Cartesian domain becomes an open-path within the radial volume. A graph-theoretic approach is utilized to find the 2D projected location of BMO points and these 2D points are projected onto the BM surface to produce the estimated location of BMO points in 3D. A machine-learning technique is employed to compute a 3D cost function around the estimated locations of BMO points. The cost function is incorporated in a dynamic programming approach to find the optimal minimum-cost path in 3D. The proposed method was compared to manual delineations (obtained by identifying the location of BMO points on 20 evenly-spaced randomly-selected B-scans manually) and the 3D iterative method proposed by Antony et al. (MICCAI 2014).

Results : Fig. 1 shows example results of BMO identification and BMO-MRW computation. The border positioning errors are reported in Table 1. The proposed method produced significantly lower unsigned and signed errors than Antony et al.’s method in the r and z directions as well as in the r–z plane (p <0.05). The proposed method, when measuring the BMO-MRW, had significantly smaller root mean square errors (RMSE) than that of Antony et al.’s method (11.62 vs. 17.99 μm).

Conclusions : The proposed method successfully identifies the 3D location of BMO points and outperformed the 3D iterative method proposed by Antony et al..

This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.

 

Fig. 1. Example results of the proposed method. The left column shows the original volumes and the column on the right shows the manual and automated BMO identification and the corresponding BMO-MRW measures.

 

Table 1. Unsigned and signed border positioning error in (Mean ± SD).