Purchase this article with an account.
N. O'Leary, M. Balasubramanian, D. F. Garway-Heath, D. P. Crabb; Developing and Testing a Heidelberg Retina Tomograph Image Simulation Method for Application to Optimise Glaucoma Progression Algorithms. Invest. Ophthalmol. Vis. Sci. 2009;50(13):2249.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Currently no gold standard exists for the measurement of the structural progression of glaucomatous neuropathy in longitudinal series of patient data. The aim was to use Heidelberg Retina Tomograph (HRT) II short-term time series data from real patients to develop a methodology to simulate realistic, virtual stable patient data with variability representative of real data.
The topographic variability in real HRT II topography images of 74 patients in a short time series study having 5 examinations within 6 weeks (Strouthidis et al Br J Ophthalmol. 2005) was measured. These data were used to seed random generation of misalignment transformations including non-linear local deformations. Variability maps smoothed by both small and large Gaussian filters were added to account for differing spatial correlations of noise. For each patient series a single topography was used from which the simulated series was propagated. To test the appropriateness of the simulated images, the mean of pixel height standard deviations (MPHSD) was compared and the correlation of pixel-wise variability was examined between simulated and real data. Rim area (RA) variability was also examined in real and simulated image sets.
Average MPHSD was 26µm for real data and 27µm for simulated data (95% CI: 10-89µm and 12-82µm respectively). The mean of cross-correlation of within patient real and simulated variability maps was 0.56 (0: uncorrelated, 1: highly correlated) with a SD of 0.15. In the images themselves, variability about certain structural features such as vessels was well reproduced whereas variability about other areas such as cup base was in some cases poorly reproduced. The mean of RA coefficient of variation (CV) values was 7% for the real data series with a SD of 7% and simulated RA CV values obtained were not significntly different (p>0.05) having mean 6% and SD 4%.
On average the simulation reproduces some of the quality of variability inherent in no-change patient HRT data. Certain areas of variability were modeled more successfully than others. This method might be useful for producing benchmark data on which competing statistical methods for detecting HRT progression can be examined for false positive rates. Future work will require refinement of the variability generation model in addition to investigating the process of simulating glaucomatous change in HRT images.
This PDF is available to Subscribers Only