Purchase this article with an account.
Jing Wu, Sebastian M Waldstein, Bianca S. Gerendas, Roland Leitner, Sinziana Birta, Georg Langs, Christian Simader, Ursula Schmidt-Erfurth; Disease Modelling & Prediction: Automated Fovea Detection as a Key Registration Landmark for Construction of a Population Reference Frame. Invest. Ophthalmol. Vis. Sci. 2015;56(7 ):5917.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To automatically compute the foveal position in multi-vendor retinal spectral domain optical coherence tomography (SD-OCT) scans of patients with exudative macular disease (neovascular age-related macular degeneration (nAMD), diabetic macular edema and retinal vein occlusion (RVO)), used as key landmarks in the construction of a population reference frame for cross-patient spatio-temporal and group-wise disease modelling and prediction.
Initially, preprocessing is performed on each OCT scan: Z dimension motion correction, denoising by block matching collaborative filtering, and graph cut segmentation to delineate the internal limiting membrane (ILM) and cysts.<br /> First, the fovea type is distinguished. Combining the segmented contiguous ILM segments generates a 3D model of the probable fovea region (Fig 1.), from which 3 distinct fovea types are examined: 1) normal foveal depression (NFD), seen as a prominent depression spanning several contiguous B-scans (Fig 1.a); 2) minor foveal depression (MFD) which features a smaller depression elevated by asymmetric retinal edema and the presence of cysts (Fig 1.b); 3) absent foveal depression type (AFD) where no depression is seen due to retinal edema by cysts (Fig 1.c).<br /> Fovea position computation for NFD (Fig 2.a) identifies the centroid of all zero thickness positions between the ILM and RNFL surfaces in the masked region. For MFD & AFD (Fig 2.b,c), pairwise distance comparison between the ILM and cyst boundary point sets is performed. From the resulting closest spatial co-ordinate pair, the position xfovea in the B-scan plane is taken as xILM, yfovea is the current B-scan and zfovea is the corresponding position on the ILM surface zILM.
Results from three disease groups each with 100 SD-OCT scans present mean (±SD) absolute distances between automated fovea detection results and manual annotated fovea for nAMD, branch RVO and central RVO to be 176.5±156.8 µm , 159.5±127.0 µm, and 165.0±143.8 µm respectively.
The presented method automatically and accurately computes the fovea position, the key landmark for creating a population reference frame, in diseased scans from “big data”. Thus this allows patient scans from different time points, vendors and modalities to be analysed in a common reference frame facilitating further advanced clinical analysis, disease modelling and prediction.
This PDF is available to Subscribers Only