Purpose:
To develop an automatic system to segment multiple retinal layers in three-dimensional spectral domain optical coherence tomography (3D SDOCT) images and to evaluate its performance in comparison with human assessment.
Methods:
Spectral rounding, a graph partitioning image segmentation algorithm, was applied to weighted 4 degree lattices constructed from 3D Cirrus OCT (Carl Zeiss Meditec, Inc., Dublin, CA) scan slices (200x200x1024 samplings in 6x6x2 mm centered on macula). Intensity differences between adjacent pixels were computed to weight the lattice. Small magnitude eigenvectors of the resulting matrix representation, the normalized laplacian of the graph, were calculated and used to determine probable boundary regions in the scan. The scans were automatically segregated into 5 regions, requiring 4 boundary detections.
Results:
The proposed method successfully segregated the scans into 5 regions (Figure). The percentage of automatically detected boundary pixels within (+/- 3 pixels) of the human specified curves are given for the 4 boundaries, color-coded in the Figure. The aggregate accuracy for the detected boundaries was 97.1% (red), 86.9% (green), 98.8% (blue), and 85.5% (yellow). These statistics were aggregated over 9 pathological subjects and 2 normative subjects - testing against hand segmentations of 4 randomly selected slices per case.
Conclusions:
The graph theoretic model tracks complex boundary contours between anatomic ocular regions. This differs from conventional segmentation algorithms such as adaptive contours, which tend to fail in the presence of retinal pathology.
Keywords: image processing • imaging/image analysis: clinical • imaging methods (CT, FA, ICG, MRI, OCT, RTA, SLO, ultrasound)