Assessing the Accuracy and Reproducibility of Retinal Segmentation on Images Acquired With HD-OCT

Abstract

Purpose:: The development of spectral-domain Optical Coherence Tomography systems has allowed the introduction of new imaging modalities, including the acquisition of three-dimensional datasets. With the help of segmentation techniques the retinal geometry can now be visualized and analyzed. The purpose of this study is to evaluate the accuracy and reproducibility of retinal thickness maps generated by a segmentation algorithm developed at Bascom Palmer Eye Institute. It should be noted that the algorithm’s performance is evaluated on the full range of pathologies presented in our patient population.

Methods:: We focus here on raster scans covering a 6x6x2 mm volume and acquired using a Zeiss HD-OCT prototype. We selected a scan pattern producing 200x200 A-scans equally spaced on the retina. Automated segmentation of the internal limiting membrane (ILM) and anterior retinal pigment epithelium (RPE) boundaries generates surfaces in 3-D space and retinal thickness maps.Eyes were grouped by diagnosis and 150 were randomly selected representing the full range of retinal diseases. Also included were 30 eyes of normal volunteers and 30 eyes of glaucoma patients. For each eye a HD-OCT dataset was entered in the study. A predetermined sample of B-scans on each dataset was manually segmented by retinal specialists and the manual segmentation was compared with the automated one. Furthermore for a subset of eyes three different scans were analyzed to assess the reproducibility of the thickness maps.

Results:: The boundaries selected from the automated algorithm are generally in very good agreement point-wise with the manually drawn boundaries. In particular on diseased eyes these lines are within 20µm of each other on well over 90% of pixels. It should be kept in mind that many of these eyes were seriously diseased and different retinal specialists were often at odds on where boundaries (especially the RPE) should be drawn.The variance of the segmentation of different scans of the same eye was measured at every pixel after the images were registered. This point-wise variance, as well as the mean variance, was typically ~ 5 µm.

Conclusions:: We show that our segmentation algorithm performs quite well on a wide mix of HD-OCT images. It is important for the clinical use of OCT that physicians be able to trust measurements such as thickness maps. This is especially true for HD-OCT, because it is difficult for the physician to evaluate directly the output of an algorithm due to the size of the datasets.

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