Abstract
Purpose: :
To develop an automatic SDOCT segmentation algorithm that provides thickness and relative locations of retinal layers. To compare results from the algorithm and human graders.
Methods: :
To segment retinal images, we developed a gradient-based, coarse-to-fine, dynamic programming algorithm. Since, some retinal layers produce stronger gradients than others, we incorporated two strategies to minimize the influence of strong boundaries on localization of nearby weak ones. First, the boundaries of the stronger gradients are used to limit the search area for those with weak gradients. Second, image gradient thresholding is used to equalize boundary contrasts. This coarse segmentation refined the search area for fine-level dynamic programming, based on the original gradients. This results in a coarse-to-fine segmentation procedure.SDOCT data, 3 B-scans each (superior, foveal, and inferior) of 20 deg were digitized with 1024 A-scans each from 5 normal subjects (Heidelberg Spectralis). The semi-automatic algorithm requires an operator to select a starting point at the left and right sides of each boundary segmented: ILM, NFL/GCL, IPL/INL, INL/OPL, OPL/ONL, ELM, IS/OS, OS/RPE, RPE/CH. We compared OCT segmentation results from the algorithm to results from 2 trained graders for 3 regions: nasal, temporal, and foveal.
Results: :
All boundaries segmented across the region of interest. Differences between the algorithm and the graders were either inconsistent across the region of interest or not found for both graders, with the exception of the ELM. The average difference between the algorithm and the graders was only 0.75 pixels. There was one small, but consistent, difference between graders and the algorithm, the ELM boundary, since graders were instructed to mark the center of the very thin ELM, and the algorithm consistently found a boundary 1.48 pixels more superficial (p < .0001, 0.004, and 0.008 for nasal, central, and temporal, respectively). A C++ software implementation took 2 min for 9 layers, vs. 12 min for manual segmentation.
Conclusions: :
The results indicate that the performance of our program is comparable with trained graders and is more efficient. The more efficient segmentation of boundaries such as those used for retinal thickness, ILM and RPE, or for monitoring retinal degeneration, IS/OS, provides the potential for larger scale normative databases for comparison with diseased eyes or biometry. Further, the small difference in ELM localization between methods allows further thickness or segmentation based on weaker gradients, e.g. ONL, by semi-automatic means that are comparable to human grading.
Keywords: image processing • imaging/image analysis: non-clinical • retina