Purchase this article with an account.
Lingjiao Pan, Xinjian Chen; Registration of 3D Retinal OCT Images. Invest. Ophthalmol. Vis. Sci. 2017;58(8):641. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Image registration has not been well-developed on 3D OCT images. The successful development of registration techniques for 3D retina OCT can greatly help us to understand the relationship between retina.
The proposed method consists of preprocessing, feature extraction and two stage registration. In preprocessing, the OCT images are segmented by ILM, NFL-GCL boundary, IPL-INL boundary, INL-OPL boundary, OPL-ONL boundary, ISL-CL boundary and BRM using multilayer graph-search approach. Then, the intensity values of layers between ISL-CL boundary and BRM are averaged to obtain the projection image. In feature extraction, edge feature is extracted from the segmented images. Edge feature takes one of eight discrete values which represents eight edge types including non-edge and seven segmented boundaries. Vessel map is extracted from the projection image using vessel enhancement filter and a mathematical morphology method. The proposed registration method is a two stage registration. The first stage is to find the displacement in x-y plane. The vessel map of subject OCT is registered to the vessel map of template OCT using SURF algorithm. The second stage is to perform non-rigid registration along z direction. It is restricted to corresponding pairs of A-scans. The hierarchical deformation mechanism of HAMMER which is successful used in MRI of brain image is adopted for the registration. The distinct anatomy such as the edge voxels on ILM and BRM are first selected to initial the deformation. Then, other edge voxels are hierarchical added to drive the deformation.
9 retinal OCT scans acquired using Topcon DRI OCT-1 scanner with image dimension of 512x992x256 and resolution of 6mmx6mmx2.5mm were included in this study. The absolute difference map between template and subject show that the registration greatly reduced their differences. The average absolute boundary surface error compare with manual segmentation is reduced from 39.9um to 7.8 um after registration.
We have presented a novel image registration approach for 3D retinal OCT image. Edge, vessel and intensity features are extracted and a two stage registration mechanism was adopt to deform the OCT image. Simulation results show the accuracy of this method.
This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.
Fig.1 Illustration of the segmentation image (a) and projection image (b).
Fig.2 (a) Template, (b) Subject,(c) Absolute difference image before registration, (d) Absolute difference image after registration.
This PDF is available to Subscribers Only