Segmentation Of Anterior Optical Coherence Tomography Images

Segmentation as part of image processing is a dividing of the digital image into connected regions, it is important and complicated task, also very crucial for feature extraction and biomedical image measurements (Deserno, 2011; Rogowska, 2006). Segmentation can be pixel-, edge-, texture- or region-oriented, and also hybrid approach can be considered (Deserno, 2011; Yang et al., 2010). In medical imaging, Canny Edge algorithm often is used to find a boundary between anatomical structures (Yang et al., 2010). The purpose of the study was to evaluate commercial retinal SD-OCT Canny Edge algorithm application in healthy corneal epithelium and endothelium boundary detection to create corneal thickness maps (Rabbani et al., 2016).

Methodological study conducted at Istad Ur & Optikk, Andalsnes, Norway in accordance with Helsinki Declaration.

On 5 eyes from 5 subjects (aged 62 – 77 years), 12-line radial corneal scans were performed using retinal SD-OCT Maestro 3D (Topcon) with anterior segment attachment. In total 60 corneal cross-sectional B-scans were generated and analysed. Maestro 3D OCT exports data to image processing and analysis by software Imagenet6 where algorithm delineated boundaries can be evaluated. 

Using MATLAB, automated segmentation CT B-scan profile (I) was compared with manually modified CT B-scan profile classified as ground truth (GT). The 6 mm  lenght radial scan pattern's each single line was divided into sectors: C – central 2 mm diameter; periphery-1 (P1+, P1-) – area from 2 mm to 5 mm diameter ; periphery-2 (P2+, P2-) – area in diameter from 5 mm to 6 mm. Data analysis was performed using MedCalc® Statistical Software version 20.106. Bland-Altman plot was calculated to compare algorithm performance of I and GT in segments – centre, periphery-1, and periphery-2. Coefficient of variation (CoV %) and inter class correlation coefficient (ICC_1,k) were used to evaluate similarities in sectoral CT measurements between I and GT.

According to Bland-Altman plot, mean difference between automated segmentation and manually modified CT measurements was 0.12 mm in centre C;  -3.33 mm in P1-, -7.25 mm in P1+, -19.80 mm in P2-, and -35.58 mm in P2+. Between I and GT, similarity of CT measurements was high at C segment (CoV = 0.33%, ICC_1,k = 0.99), good at periphery-1: P1- (CoV = 0.91%, ICC_1,k = 0.98) and P1+ (CoV = 1.43%, ICC_1,k = 0.96), but average at periphery-2: P2- (CoV = 3.59%, ICC_1,k = 0.79) and P2+ (CoV = 5.37%, ICC_1,k = 0.55).

A Canny Edge algorithm implemented in Maestro 3D SD-OCT performs excellent corneal epithelial and endothelial segmentation in corneal central area, but performance is insufficient at the peripheral area in diameter from 2 to 6 mm. Therefore, only central corneal thickness measurements can be considered for diagnostic purposes. In periphery, the algorithm underestimates corneal thickness. Therefore, this area can be overviewed only as reference and manual adjustments are required.