Segmentation of Three-dimensional Retinal Image Data

Abstract

Abstract-We have combined methods from volume visualization and data analysis to support better diagnosis and treatment of human retinal diseases. Many diseases can be identified by abnormalities in the thicknesses of various retinal layers captured using optical coherence tomography (OCT). We used a support vector machine (SVM) to perform semi-automatic segmentation of retinal layers for subsequent analysis including a comparison of layer thicknesses to known healthy parameters. We have extended and generalized an older SVM approach to support better performance in a clinical setting through performance enhancements and graceful handling of inherent noise in OCT data by considering statistical characteristics at multiple levels of resolution. The addition of the multi-resolution hierarchy extends the SVM to have "global awareness." A feature, such as a retinal layer, can therefore be modeled within the SVM as a combination of statistical characteristics across all levels; thus capturing high-and low-frequency information. We have compared our semi-automatically generated segmentations to manually segmented layers for verification purposes. Our main goals were to provide a tool that could (i) be used in a clinical setting; (ii) operate on noisy OCT data; and (iii) isolate individual or multiple retinal layers in both healthy and disease cases that contain structural deformities.

A Review of Algorithms for Segmentation of Retinal Image Data Using Optical Coherence Tomography

“…In addition, an alternative promising method was introduced by Mayer et al, who used a fuzzy C-means clustering technique to automatically segment RNFL thickness in circular OCT B-scans without the need of parameter adaptation for pathological data (Mayer et al, 2008). In contrast to the edge detection approaches mentioned above, a multi-resolution hierarchical support vector machine (SVM) was used in a semi-automatic approach to calculate the thickness of the retina and the photoreceptor layer along with the volume of pockets of fluid in 3D OCT data (Fuller et al, 2007). In this approach, the SVM included scalar intensity, gradient, spatial location, mean of the neighbors, and variance.…”

“…This methodology also included a semi-supervised approach to correct for segmentation errors such as false regions marked as drusen in images showing RPE elevation unrelated to drusen. The approach presented by Fuller et al and described above also facilitates the semi-automatic segmentation of drusen in SDOCT images (Fuller et al, 2007). Gregrori et al has also measured drusen area and volume using quantitative descriptors of drusen geometry in three dimensional space (Gregori et al, 2008).…”

Ocular fundus imaging: From structure to function

“…Several segmentation methods have been used in OCT, such as gradient [10,11] and support vector machines (SVM) based methods [6,7] among others. While gradient methods usually do not work in areas of low quality signal, SVM methods require a previously segmented training set.…”

“…We stress that there is no need of a manual segmentation to establish a training set and that the number of used features is small considering published literature on this field [6,7].…”

“…The classification for the training set is obtained by gradient methods, namely the mean-shift algorithm. No manual input is needed, opposite to other segmentation methods based on classification [6,7].…”