Purchase this article with an account.
omar helmy, Ahmed ElTanboly2, Ahmed Soliman, Ahmed Sleman, Harpal Sandhu, Ayman El-Baz, Shlomit Schaal; Automated 3-D segmentation of retinal layers from optical coherence tomography images using pattern recognition-based method. Invest. Ophthalmol. Vis. Sci. 2018;59(9):1685. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
The purpose of this study is to develop a robust automated approach capable of 2-D/3-D segmentation of retinal layers from (OCT) images. OCT image segmentation is applied to identify the intra-retinal boundaries between the retinal layers. The ultimate goal of the proposed segmentation method is to improve the accuracy of this process and to reduce required processing time. Current OCT machines adopt technologies that are capable of capturing the volumetric part of the retina (3-D volumes), therefore are now moving towards volume segmentation along with 3-D rendering and visualization.
SD-OCT scans (Zeiss Cirrus HD-OCT 5000) were prospectively collected from ten normal subjects (age 10-79) without pathology. Each 3-D scanned volume was sliced into five 2-D images, patients underwent macular scans that were captured at ~840nm wavelength and with a scan speed of 68,000 A-scans per second. OCT image segmentation is comprised of two steps and performed as follows: in the first step, the centered image wsegmented into 12 layers using a joint model that integrates shape, intensity, and spatial information and consists of three basic sub-steps: (i) aligning the input centered image to the constructed shape database, which consists of different central images collected from different healthy and diseased subjects, (ii) applying the joint model to the aligned image, and (iii) obtaining final segmentation.The second step is based on utilizing the obtained central image segmentation to guide the segmentation process for other neighboring images (inside the 3-D volume)in a progressive way. In this process, a shape prior which is based on the image's gray and labeled vallues is used as an initial point of segmentation. The Shape prior is then utilized to guide segmentation of the adjacent images.
The segmentation accuracy has been evaluated in terms of Dice similarity coefficient (DSC). An average DSC of 82.6 ± 3.34% has been obtained over the ten used scans. To highlight the accuracy of the proposed approach, the segmentation results were compared to one of the state of the art methods  that has achieved an average of 78.7 ± 2.92%.
The introduced segmentation algorithm carries promise for 3-D visualization in addition to extracting more retinal layers characteristics that could help in diagnosing retinal diseases.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.
This PDF is available to Subscribers Only