Texture Analysis and Geometry Based Retinal Segmentation for Three-Dimensional OCT

Abstract

Purpose: : Segmentation of retinal OCT tomograms utilizing texture and geometric parameters is performed to allow for unsupervised extraction of retinal layer morphology in large three-dimensional data sets for objective quantitative diagnosis and therapy monitoring of retinal pathologies.

Methods: : Analysis is performed utilizing more than 230 different features, subsequently selected by an sequential forward search. For training, OCT images are segmented by hand and a number of texture features are extracted. A sequential forward search algorithm is used to reduce the search space to find the near best feature set, and for each layer the distribution of selected features is modeled by a Gaussian. To classify each layer Mahanalobis distance is used to determine which layer each pixel is closest to in the feature space. In the next stage, force fields are constructed reflecting the texture distribution in the image and the level set method is run for each layer, so that expanding curves mark smooth boundaries between the layers.

Results: : A set of ten normal subjects was used to build the classification set. Training data was manually segmented by independent, skilled operators and was compared to automatically segmented test data. All major intraretinal layers and their interfaces were indicated; these include the nerve fiber layer (NFL), ganglion cell layer and inner plexiform layer (GC/IPL), Henle fiber layer and outer plexiform layer (HF/OPL), outer nuclear layer (ONL), inner segment of photoreceptor layer (IS PR), outer segment of photoreceptor layer (OS PR) and the retinal pigment epithelium (RPE). The automated segmentation of these layers for the parafoveal tomogram and the central foveal tomogram was achieved with high accuracy compared to manual expert segmentation.

Conclusions: : The preliminary results already show the huge potential of this technique and are likely to improve with further training and fine tuning. Combining texture based analysis with level set method gives segmentation results superior to any OCT segmentation method presented so far.

Keywords: image processing • imaging methods (CT, FA, ICG, MRI, OCT, RTA, SLO, ultrasound)