Abstract
Purpose:
To develop a computer-aided manual segmentation method (EdgeSelect) for SD-OCT images, and compare its segmented results with those using a traditional manual segmentation (Manual) method.
Methods:
3D Spectralis SD-OCT images (Heidelberg Inc., Germany) covering the central macular region of 22 eyes (a subset of the DRCR.net subjects) were segmented by three graders for inner limiting membrane (ILM) and retinal pigment epithelium (RPE) boundaries using both the EdgeSelect and the Manual methods. The EdgeSelect method incorporates a computational step of outlining potential edges based on the local intensity gradient of the OCT image, followed by an interactive manual step for graders to confirm and adjust the selected edges, and the computer to generate the desired boundary by connecting edges using Dijkstra's shortest length algorithm. The Manual method requires graders to mark multiple points along the desired boundary, and the resultant boundary is computed using spline interpolation [Hood et.al. Optom Vis Sci 88(1)113-123]. The inter-grader reproducibility is evaluated by computing the mean value of the absolute distance at each pixel (mean Δ) between grader-pairs.
Results:
The inter-grader mean Δ for EdgeSelect and Manual methods are 0.34 μm and 3.51 μm for ILM segmentation, and 1.26 μm and 7.72 μm for RPE segmentation, respectively. Paired t-test showed that the difference between the two methods is statistically significant (p<0.001). Additionally, the EdgeSelect method reduces the segmenting time comparing to the Manual method.
Conclusions:
The EdgeSelect method yields better inter-grader reproducibility than the Manual method for segmenting OCT layers.
Keywords: 550 imaging/image analysis: clinical •
552 imaging methods (CT, FA, ICG, MRI, OCT, RTA, SLO, ultrasound) •
688 retina