Purchase this article with an account.
S. R. Russell, M. D. Abramoff, V, M. D. Radosevich, E. Heffron, E. M. Stone, E. S. Barriga, B. Davis, P. Soliz; Quantitative Assessment of Retinal Image Quality Compared to Subjective Determination. Invest. Ophthalmol. Vis. Sci. 2007;48(13):2607.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To demonstrate an automatic method for assessing retinal image quality based on simple calculated image statistics.
A set of twenty fundus images which ranged in visible image quality and severity of macular degeneration, acquired on a Zeiss FF4 30-degree camera and recorded on Ektachrome 35 mm slides were digitized using two digital scanners, a Nikon Coolscan model 4000, and a custom high-speed CCD-based digitizing system. Images were scanned at maximum spatial resolutions (5781 by 3945 pixels, 8 bits/channel for the Nikon and 3456 by 2298 pixels, 16 bit/channel for the high-speed CCD device). For each RGB channel, statistics were calculated including mean channel intensity, variance, skewness, contrast, spatial frequency kurtosis and mean red/green and blue/green ratios. A partial least squares (PLS) model was developed to determine those image features which provided the greatest power in determining image quality. For comparison, subjective image quality was graded for each image by an experienced retinal investigator, which served as a ground truth standard. Image quality was assigned as high, medium, or low, based on the clarity of image features such as retinal vessels, and on the grader's confidence in retinal diagnoses based upon the image.
The model showed that the quantitative statistics could separate the three classes of image quality. Images from both scanners successfully classified 100% of the images based on the visual classification by the human graders. Based upon the magnitude of regression coefficients for each statistic, seven features from the combined RGB fundus color images provided the necessary descrimination to categorize the images by subjective classification grade.
We have demonstrated the ability to classify retinal image quality based on simple image features that can be calculated in real-time. Further efforts to provide quantitative real-time image quality assessments could provide prompt feedback to photographers and investigators seeking the highest quality images for clinical or study use.
This PDF is available to Subscribers Only