Purchase this article with an account.
Georgios Lazaridis, Jibran Mohamed-Noriega, Marco Lorenzi, Sebastien Ourselin, David F Garway-Heath; Imaging outcomes in the UK Glaucoma Treatment Study (UKGTS): improving the statistical power of the UKGTS by OCT image enhancement via Bayesian fusion of ensemble generative adversarial networks. Invest. Ophthalmol. Vis. Sci. 2020;61(7):874.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
In trials with a vision function outcome, variability in measurements results in the requirement for large numbers of patients observed over long intervals. Image-related anatomical measurements have considerable potential as an additional outcome to improve trial statistical power. Although spectral-domain OCT (SDOCT) is now widespread, past trials used time-domain OCT (TDOCT), which leads to poor statistical power due to low signal-to-noise ratio. Improving the quality of images has great potential for increasing statistical power in clinical trials.
We propose a) a probabilistic ensemble model and b) a cycle-consistent perceptual loss for improving the statistical power of imaging-based clinical trials. TDOCT are converted to synthesized SDOCT images and segmented via Bayesian fusion of an ensemble of cyclical generative adversarial networks (GANs). We train two networks based on the proposed ensemble method with: i) GAN loss and ii) GAN+perceptual loss. For training and validation, independent datasets were used: 148 eyes lead to 24,792 TDOCT and SDOCT pairs. Testing experiments on the UKGTS dataset (373 participants with 78,415 TDOCT) are performed by quantifying the trial statistical power of measurements derived from the synthesised SDOCT images compared with those derived from the original TDOCT images.
We demonstrate that the proposed methodology for TDOCT image improvement i) significantly improves the agreement of segmented RNFL thickness measurements with SDOCT measurements and ii) significantly reduces the test-retest variability (Table 1). When the rate of RNFL loss in the UKGTS data set is calculated from the synthesised SDOCT images, the RNFL measurements are able to distinguish the difference in slopes between treatment groups (p = 0.0025) whereas TDOCT measurements are not (Table 2).
The results show a significantly better separation between treatment arms with segmentation of enhanced images than conventional TDOCT. This is the first demonstration that imaging outcomes alone can distinguish treatment effects in a clinical trial of glaucoma therapy.
This is a 2020 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only