3D Image Registration Based Atlas for Large Field of View Human Eye Optical Coherence Tomography

Abstract

Purpose : Optical coherence tomography (OCT) is a non-invasive imaging modality used for high resolution (10 μ) imaging of ocular tissue. Retinal cell layers are clearly visible on OCT providing a means for quantitative and topographical analysis. Medical atlases are important tools for detecting abnormal structural and anatomical tissue alterations, relative to a standardized normal. They can also serve as a reference for registration of individual images for deformation-based analysis. The purpose of the current study is to present a method for generating an atlas of the human eye using large field of view 3D OCT image volumes.

Methods : Imaging was performed using a commercial instrument (Spectralis, Heidelberg Engineering, Germany) in both eyes of a healthy (N=1) and subjects with non-arteritic anterior ischemic optic neuropathy (N=2), compressive optic neuropathy (N=1), Neuromyelitis optica (N=1) and cortical stroke (N=2). Images were centered on the macula and acquired in both horizontal and vertical directions. Each volume covered an area of 5.80 x 2.82 x 5.88 mm of retinal tissue and consisted of 1536 x 496 x 97 voxels. Image registration was performed separately for vertical and horizontal scans using the Elastix framework to align retinal anatomy. In each group, a volume that was not tilted and provided clear visualization of layer boundaries was selected as fixed. The remaining volumes were considered to be moving and aligned with the fixed one by B-spline transformation. The pipeline consisted of an initial affine to define the region of interest (ROI) and a non-linear registration to align layer boundaries and anatomical details. Normalized mutual information (NMI) cost function with gradient descent optimization was used to obtain an efficient registration. Quality of registration was validated by a human observer and volumes were averaged to provide the final atlas.

Results : A final volumetric atlas of a wide field of view of horizontal and vertical OCT image volumes are shown in Figure 1a and Figure 1b, respectively. As shown, retinal layers and the foveal pit have smooth boundaries which indicates high quality image registration.

Conclusions : A method for generating an atlas of retinal OCT images in large a field of view is presented. The atlas can serve as a reference for deformation-based anomaly detection in the eye and for mapping retinal cell layers to the visual cortex.

This is a 2020 Imaging in the Eye Conference abstract.