Automated Segmentation of the Retinal Vasculature Silhouettes in Isotropic 3D Optical Coherence Tomography Scans

Purpose: To report and evaluate an algorithm for the segmentation of retinal vessels in isotropic 3D spectral OCT scans, with closely spaced slices. Its performance is compared with two expert segmentations of OCT projection images and a single expert segmentation of the corresponding color fundus photographs. Methods: 16 optic nerve head centered 3D OCT volumes (200x200x1024 voxels, 6x6x2 mm3) were obtained from 8 subjects (1 scan per eye) using a Cirrus HD-OCT machine (Carl Zeiss Meditec, Inc., Dublin, CA). Additionally, 16 color disc photographs, one from each of the eyes of these 8 subjects, were acquired. Two experts manually marked the retinal vessel silhouettes in projection images derived from the SD OCT scans. The two expert's results were combined into a single OCT reference segmentation. A single retinal expert marked all vessels in the color photographs. The automatic vessel segmentation algorithm (Niemeijer et al., ARVO 2008) was applied all OCT projection images. The average segmented vessel lengths (mm) identified, missed and spuriously identified by the algorithm and the OCT reference segmentation were compared, after co-registration, with the expert detected vessels in the color photographs. Results: The algorithm, on average, correctly identified 37.5mm (95% CI, 34.7-40.3), missed 19.9mm (95% CI, 17.4-22.4) and spuriously identified 4.8mm (95% CI, 3.7-5.9) of vessels, compared to the color fundus photographs based standard. The OCT reference segmentation identified 39.7mm (95% CI, 36.4-43.0), missed 17.6mm (95% CI, 15.4-19.8) and spuriously identified 2.0mm (95% CI, 1.6-2.4) of vessels. Conclusions: The automatic system detects close to the same length of vessels in OCT projection images compared to an expert. The length of the detectable vasculature is reduced in 3D spectral OCT compared with fundus images. Given this performance, automated vessel segmentation in isotropic SD OCT can help co-localize OCT scans to each other as well as to fundus photos from the same patient at different timepoints.