Purchase this article with an account.
S. Murillo, G. Zamora, S. Nemeth, R. Crammer, A. Edwards, W. Bauman, P. Soliz, M. Pattichis; Low-Cost Super Resolution Retinal Imaging With Embedded Denoising: Quantitative and Qualitative Assessment of Reconstructed Images From a Scanning Laser Ophthalmoscope. Invest. Ophthalmol. Vis. Sci. 2010;51(13):1803.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To assess the applicability of a Super Resolution (SR) algorithm for increasing information content while enhancing the visualization of physiologically significant landmarks by generating SR images of the retina from multiple low resolution pictures of the same scene.
The dataset consisted of 32 retina video sequences from normal subjects taken with a low-cost SLO. Each video consisted of 20 frames; each of size 1024x1024 pixels with a footprint of 10 um and acquired at 10 frames per second. Regions of interest (512x512 pixels) from 4 to 16 frames were used to test a SR algorithm that uses a Total Variation (TV) functional to increase image resolution by factors of 2X to 4X. The functional contains an embedded denoising term that corrects for acquisition artifacts introduced by the SLO. Quantitative metrics of information content and image quality were applied to the original and SR images to assess changes introduced by the algorithm. Qualitative assessment of utility in discerning four physiologically salient features on the retina (disc margins, major and minor vessels, and macula) was performed by three highly experienced individuals in a masked, pair-wise comparison experiment of the original image to the SR version. Inter-grader agreement (kappa) in the choice of preferred image was calculated.
There was no statistically significant difference in the quantitative metrics between the original and SR images (p=0.01), i.e. all original and SR image pairs had essentially the same contrast, brightness, and entropy. In the qualitative assessment, all graders chose the SR images over their original counterparts in all cases resulting in perfect inter-grader agreement (kappa=0.1).
The video frames contributed with sub-pixel information to reconstruct SR images. The quantitative metrics remained unchanged indicating that SR does not create information. Rather it combines the information of all the observations into a single image of increased pixel detail aiding the clinician into making a better evaluation of the retinal features. The results demonstrate that more useful images can be obtained without incurring in higher acquisition costs through an efficient algorithm that combines the non-redundant information contained in images of the same scene.
This PDF is available to Subscribers Only