Purchase this article with an account.
Alessio Montuoro, Sebastian M Waldstein, Bianca Gerendas, Georg Langs, Christian Simader, Ursula Schmidt-Erfurth; Statistical Retinal OCT Appearance Models. Invest. Ophthalmol. Vis. Sci. 2014;55(13):4808.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To generate a noise and motion artifact free synthetic Optical Coherence Tomography (OCT) data set as a standardized ground truth against which image processing algorithms can be tested and validated.
More than 20.000 OCT scans of over 1000 patients retrieved from the Vienna Reading Center (VRC) database were used as a basis for the generation of the synthetic data set. The scans were grouped by vendor (Spectralis OCT, Heidelberg Engineering and Cirrus HD-OCT, Carl Zeiss Meditec) and for each scan the internal limiting membrane (ILM) and retinal pigment epithelium (RPE) segmentation of the device specific software where extracted. The thickness profile (i.e. distance between ILM and RPE) was calculated and regions with missing or implausible thickness values were excluded automatically. After estimating the foveal position, the polar coordinates relative to the macular center were calculated for each A-scan. By grouping spatially close A-scans with same thickness values, an appearance model for each position and thickness was estimated. A mathematical model of the thickness map was constructed (Fig. 1) and used to select corresponding A-scans out of the appearance model. The A-scans of the final synthetic volume were interpolated using a weighted circular neighborhood in order to avoid artifacts introduced due to the spatial grouping of the source A-scans.
The method yields a synthetic, noise and motion artifact free representation of healthy OCT scans. It enables generation of multiple OCT volumes of varying scan geometries and positions relative to the macula. After adding synthetic speckle noise (Fig. 2) a realistic volume is derived.
The availability of an OCT ground truth simplifies the comparison of different classes of image processing methods such as denoising, segmentation or motion correction algorithms. By generating OCT volumes based on different OCT devices using the same scan parameters, cross-device comparison of the repeatability of image processing algorithms is feasible. We are currently implementing the method for additional OCT devices and adding additional retinal features like vessel shadows and motion artifacts.
This PDF is available to Subscribers Only