Abstract
Purpose: :
We have previously published a graph-theoretic approach that simultaneously segments multiple intra-retinal surfaces at the optic nerve head (ONH) in 3-D from volumetric OCT scans (Antony et al., SPIE 2010). Traditionally, cost functions were designed by hand, but in the case of diseased scans a careful design may prove beneficial. Furthermore, in the presence of pathology induced variation in layer thickness and appearance, texture may supplement traditional image descriptors. Here, we present a method for the designing of cost functions that incorporate texture features learned from a training set.
Methods: :
The intra-retinal layers were grouped by intensity, and probability maps (which reflect the likelihood of a voxel belonging to a particular region) were created for the groups by classifying Gabor texture features. Alternatively, for the purposes of comparison, probability maps were also created without grouping the layers prior to classification. These probability maps were then incorporated into the cost functions utilized by the graph-theoretic approach, and used to segment the intra-retinal surfaces from 10 ONH centered OCT volumes obtained from 10 subjects that presented with glaucoma. The mean unsigned border position error was compared to our previously published results as well as the alternative approach.
Results: :
The overall mean segmentation errors noted after the incorporation of texture was found to be 6.9 ± 2.4μm, which is significantly smaller (p < 0.01) than our previously reported result 8.94 ± 3.8μm, as well as the alternative approach 9.08 ± 4.4μm. The mean is also comparable to the inter-observer variability 8.52 ± 3.6μm.
Conclusions: :
Texture, which finds little use in current approaches, is an important descriptor of the retinal layers seen in OCT volumes and the incorporation of these learned features into the optimal graph-theoretic approach can significantly improve the accuracy of the segmentation of intra-retinal surfaces. Furthermore, the grouping of regions based on similar properties can increase the accuracy of the classification, and by extension the layer segmentation.
Keywords: image processing • texture • imaging methods (CT, FA, ICG, MRI, OCT, RTA, SLO, ultrasound)