Purchase this article with an account.
Xinyu Song, Rolando Estrada, Stephanie J. Chiu, Al-Hafeez Dhalla, Cynthia A. Toth, Joseph A. Izatt, Sina Farsiu; Segmentation-Based Registration of Retinal Optical Coherence Tomography Images with Pathology. Invest. Ophthalmol. Vis. Sci. 2011;52(14):1309.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Accurate construction of 3D retinal volumes from SDOCT is hindered by the presence of patient eye motion. Additionally, common volume acquisition protocols result in low azimuthal resolution along the direction of successive B-scans. To improve the accuracy of quantitative volumetric measurements, e.g. drusen volume, it is desirable to attain isotropic lateral resolution by capturing and fusing the information of 2 volume scans at 0- and 90- degree orientations. Unfortunately, the fundamental assumption that registered frames are inherently similar limits application of most classic image registration techniques for this task. We have previously developed algorithms for reliable segmentation of up to 8 intra-retinal boundaries. In this project, we utilize the segmentation information to aid in reconstructing SDOCT volumes free of motion artifacts.
We captured 2 perpendicularly oriented SDOCT volumes with 100 B-scans (1000 A-scans each) from a 6.6mm square retinal area of five subjects with non-neovascular AMD. Many of the B-scans included relatively large drusen. We used our software to segment the retinal layers. To register consecutive B-scans, we minimized the rate of change in location of these layers along the azimuthal direction. Moreover, we adjusted the position of B-scans in each volume by enforcing a retinal anatomy condition estimated by segmenting the corresponding layers in the perpendicularly oriented volume. To register the two perpendicularly oriented SDOCT volumes, we detected and matched the vessels on their axially projected summed-voxel-projections. As a quantitative measure of registration reliability, we defined the rate of the nerve fiber layer (NFL) position change (RNPC) as the absolute value difference between the pixel position of the NFL in one frame and the average position of the NFL in the before and after frames.
Visual examination of the registered 3D volumes attested to the accuracy of the proposed technique. The RNPC mean and standard deviation at 0- and 90- degree volumes were 3.0 ± 1.2 and 2.9 ± 1.1 pixels for non-registered frames and 1.8 ± 0.7, 1.7 ± 0.5 pixels for registered frames, respectively.
We achieved reliable registration of volumetric B-scan data through segmentation of B-scans in combination with a small number of B-scans acquired in the orthogonal direction.
This PDF is available to Subscribers Only