Purchase this article with an account.
Meindert Niemeijer, Mona K. Garvin, Zhihong Hu, Kyungmoo Lee, Milan Sonka, Michael D. Abramoff; Automated Registration Of Fundus Photographs To 3D OCT Projection Images. Invest. Ophthalmol. Vis. Sci. 2011;52(14):1304.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
A method to automatically register 3D OCT projection images with fundus photographs is reported. Registration is essential for aligning these images in image guided therapy as well as to compare images acquired on different machines at different times.
For this work, 19 deidentified ONH-centered SD-OCT scans (Zeiss Cirrus HD-OCT) and corresponding stereo color fundus photographs (Nidek 3Dx) from the same 19 subjects with glaucoma were acquired at the University of Iowa Hospitals and Clinics. Each SD-OCT scan consisted of 200 × 200 × 1024 voxels, while each stereo color fundus photograph was 4096 × 4096 pixels. The retinal layers in each SD-OCT scan were segmented and a vessel silhouette projection image was produced. This 2D projection image was registered with the fundus photograph. In both the fundus photograph as well as the projection image an automated vessel segmentation method was applied. Our method assigns each pixel in the image a likelihood that the pixel is inside a vessel. The Scale Invariant Feature Transform (SIFT), a general feature-point detector, was applied to the vessel maps. For each pair of fundus and projection images the detected feature-points were compared and matched. The set of matching featurepoints were further processed with the RANdom SAmple Consensus (RANSAC) algorithm to remove outliers. After RANSAC, between 10 and 20 matching featurepoints remained. Based on this set of points the similarity registration transformation that minimized the total error was found.
A human observer was asked to select a single point in each of the image pairs. After registration of the fundus images, errors in the OCT images were assessed. The average error was 1.89 pixels(56.7µm) with a standard deviation of 1.15 pixels(34.5µm).
Feature-point based registration works well for the registration of fundus images with OCT projection images. The resulting registration error is small.
This PDF is available to Subscribers Only