Purchase this article with an account.
Lennart Husvogt, Stefan Ploner, Eric M Moult, A. Yasin Alibhai, Julia Schottenhamml, Jay S Duker, Nadia K Waheed, James G Fujimoto, Andreas K Maier; Using Medical Image Reconstruction Methods for Denoising of OCTA Data. Invest. Ophthalmol. Vis. Sci. 2019;60(9):3096.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
A commonly used method to generate OCT angiography (OCTA) data is to compute the amplitude decorrelation of repeated B-scans. Despite its prevalence, to our knowledge, amplitude decorrelation, and related metrics were developed heuristically, and lack complete theoretical descriptions. Outside of OCTA, a variety of compressed sensing-based image reconstruction algorithms have been successfully applied to magnetic resonance imaging and computed tomography. Inspired by the work in these fields, we developed a probabilistic model for amplitude decorrelation. This, and models for speckle variance and interframe variance, enable an objective-function minimization approach to OCTA data generation with optimized noise characteristics.
We generated ground-truth images by registering and merging 10 consecutively acquired 3x3mm OCTA volumes from a healthy volunteer. A compressed-sensing-based denoising method with a 3D median filter for regularization was used for reconstruction.
Figure 1 shows the decreasing mean squared error, compared to our ground-truth data, of our reconstruction algorithm over 100 iterations, indicating increasingly improved noise characteristics. Figure 2 shows corresponding representative en face retinal OCTA images from the reconstruction; ground truth data are shown in panel A, and test data in panel B. Compared to median filtering (panel C), our OCTA reconstruction decreases noise while minimizing image blurring. Reconstruction results in panels D through F show how the reconstruction can be used to optimize the denoising between the original volume and the median-filtered volume.
State-of-the-art reconstruction techniques, such as compressed sensing, can be adopted from other medical imaging fields to improve the quality of OCTA data.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
The mean squared error between the ground truth and the reconstruction result is shown in blue. The mean squared error between the ground truth and the input volume after applying a median filter is shown in green.
3x3mm en face speckle variance projections of a 28 year old normal person. From top left to lower right: a: ground truth data b: input data before reconstruction, c: input data after median filtering, d: reconstruction after 10 iterations, e: 20 iterations, d: 30 iterations
This PDF is available to Subscribers Only