Purchase this article with an account.
Qingyang Su, Jui-Kai Wang, Mohammad Saleh Miri, Victor A. Robles, Mona K Garvin; Spectral-Domain Optical Coherence Tomography Optic-Nerve-Head and Macular En-Face Image Registration in Cases of Papilledema. Invest. Ophthalmol. Vis. Sci. 2017;58(8):646.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Previously, we proposed an automated method to register fundus images with its corresponding spectral-domain optical coherence tomography (SD-OCT) retinal pigment epithelium (RPE) en-face images in glaucoma (Miri et al. BOEx 2016). In this work, we extended the method to register two SD-OCT RPE en-face images from the optic-nerve-head (ONH) and macular scans in cases of papilledema. This work overcame two current challenges: 1) very limited overlap region between the two en-face images, and 2) frequent appearance of the massive image shadow around the ONH due to papilledema.
The proposed algorithm, first, searched for corners using the features from the accelerated segment test (FAST) approach in both ONH and macular en-face images. Next, the histograms of oriented gradient (HOG) for each selected corner were computed, and the proposed algorithm decided potential mapping landmarks by identifying the best matches of the feature descriptors (Fig. 1). Then, the proposed algorithm removed the incorrect landmark pairs based on the geometrical distribution of all candidates. Finally, the proposed algorithm utilized random sample consensus (RANSAC) method to estimate a similarity transformation matrix and generated the registered panorama image (Fig. 2).
Fifty subjects were randomly selected from the Idiopathic Intracranial Hypertension Treatment Trial baseline dataset - where 30/20 subjects with 57/39 available OCT ONH and macular image pairs were in the training/testing set, respectively. The parameters in the proposed algorithm were empirically determined using the training set. In the testing set, two manual landmark pairs for each OCT image pair were first selected between the fixed and moving images. To evaluate the registration results, the mean distance of these landmarks shifted by the proposed algorithm in the moving image was calculated. Overall, the mean unsigned difference in the testing set was 1.97±1.00 pixels (59.1±30 µm).
With accurate image registration, quantitative measurements at the region between the ONH and fovea (such as retinal layer thicknesses in grids, wide-field-retina-shape measures, retinal fold quantification) are potentially clinically meaningful in papilledema.
This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.
Fig. 1. Examples of HOG Descriptor Computation
Fig. 2. Example of Landmark Selections and Registration Results
This PDF is available to Subscribers Only