Performance Of The Sd-Oct Image Segmentation Algorithm For Healthy And Corneas With Surface Irregularities

The purpose of study was to evaluated the performance of the commercial SD-OCT Maestro 3-D (Topcon) image processing software Imagnet6 Canny Edge segmentation algorithm in determining the boundary surfaces of the corneal layers in healthy corneas and in ectatic corneas.

Cross-sectional corneal B-scan thickness profiles generated by Canny Edge automatic segmentation (I) were compared with manually modified segmentation (GT) transformed B-scan thickness profiles. A total of 60 B-scans of healthy corneas and 60 B-scans of ectactic corneas were analyzed. Bland-Altman plot used to analyse automatic segmentation performance in healthy corneas at central 2 mm area, paracentral 2 - 5 mm area (P1), and paracentral 5 - 6 mm area (P2). Ectactic corneas were analysed in the step of 1 mm rings.

In the healthy corneas automated segmentation algorithm slightly overestimated corneal thickness (CT) in the central zone (mean 0.12 μm) and significantly underestimated (p < 0.05) CT at zones P1 (mean -3.33 μm and -7.25 μm) and P2 (mean -19.80 μm and -35.58 μm). In ectatic corneas automated segmentation algorithm significantly overestimates (p < 0.05) CT in central area (mean 8.42 μm) and underestimates CT in paracentral zones (mean from -9.70 μm to -86.67 μm). There were no significant differences between automated and manually modified segmentation in central area 1 mm to 2 mm. The mean difference according Bland-Altman plot in C1+ zone was 4.37 μm (p > 0.05) and in C1- zone 4.22 μm (p > 0.05).

For healthy corneas, the Canny Edge algorithm provides excellent segmentation of the corneal epithelium and endothelium in the central zone, but the performance of the segmentation algorithm is insufficient in the mid-peripheral zone between 2 mm and 6 mm in diameter, since in this zone the automatic segmentation algorithm detects an erroneously lower corneal thickness. For ectatic corneas, segmentation algorithm in the Maestro 3D (Topcon) device fails to provide a sufficiently good segmentation of the corneal epithelial-endothelial boundary surfaces in the center and periphery. The algorithm has a tendency to determine the corneal thickness erroneously higher in the center, and erroneously lower in the paracentral sectors.