Abstract
Purpose :
To find practical statistical models of noise in individual retinal layers of images from commercial spectral-domain optical coherence tomography (SDOCT) systems and to attain a bound on accuracy of retinal layer segmentation techniques.
Methods :
The Cramer-Rao lower bound (CRLB) determines the minimum possible variance of an estimate of a parameter, such as the location of retinal layer boundaries in OCT images. Calculation of the CRLB requires models of the relevant noise distributions, but empirically verified models of layer-specific noise in commercial SDOCT systems, which may contain proprietary image processing algorithms, are not available. Using data from a Bioptigen SDOCT system, we determined noise models specific to 9 regions of retinal SDOCT B-scans and calculated the CRLBs for the 8 boundaries between those regions. Two expert manual graders segmented 55 intensity-corrected parafoveal B-scans from 5 patients imaged with a Bioptigen SDOCT system into 9 regions as shown in Fig. 1. We created 9 region-specific intensity histograms from all pixels belonging a given region across all images. Each histogram was fit to 8 literature-derived noise distributions, and accepted or rejected based on the goodness of fit. The accepted fit with the fewest free parameters was considered the best noise model. Using the best noise models for the regions adjacent to each of the 8 boundaries, we calculated the CRLBs. We also calculated optimally-biased CRLBs, which take into account the prior information often used by segmentation algorithms.
Results :
Each region’s histogram fit successfully to at least one distribution, but no single distribution fit to every region. The unbiased CRLBs were calculated to be roughly two orders of magnitude higher than the biased CRLBs.
Conclusions :
Our results show that it is possible to independently model the noise of multiple regions of the retina, and that these models can be used to calculate the bounds of segmentation accuracy. The biased lower bounds are substantially lower than the variance of current layer segmentation algorithms, indicating the possibility of improvement. Although these results are specific to a particular commercial SDOCT system, the technique can be applied to any commercial OCT system.
This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.