Purchase this article with an account.
P. Soliz, B. Raman, S. Nemeth, S. Barriga, G. Zamora, E. Bursell; Improved Automated Digital Retinal Image Analysis for Detection of Diabetic Retinopathy Through Image Quality Restoration . Invest. Ophthalmol. Vis. Sci. 2005;46(13):359.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Purpose: Image quality greatly affects the effectiveness with which human graders and computer–based systems can perform diabetic retinopathy (DR) screening using digital retinal images. The aim of this study was to demonstrate the significance of two critical methodologies for improving technical image quality after image acquisition on the sensitivity and specificity for detecting retinopathy in diabetic patients. Methods: A total of N= 42 subjects (N=74 eyes, 33% with no diabetic retinopathy (NDR) and 67% with diabetic retinopathy (DR)) were processed. One set of data were provided by the Joslin Vision Network system which uses a non–mydriatic digital color fundus camera. A second set was collected through a dilated pupil. This algorithm is based on Marr’s vision model for edge detection and its tolerance for significant illumination differences, as in shadows. Additionally, image blur due to ocular aberrations or camera de–focus were corrected analytically. Segmentation of the micro–aneurysms was performed. A trained retinal grader determined which of the segmented objects were either true positives or false negatives, i.e. ground truth. The segmentation process was performed on the un–processed and the enhanced images. A comparison of the clinical sensitivity and specificity was performed for the same classification algorithm when using the unprocessed and enhanced images. Results: The MA segmentation algorithm had a sensitivity and specificity of 69% and 66% when applied to the unprocessed non–mydriatic images. After image enhancement, the sensitivity and specificity improved to 98 and 75%. For the mydriatic images, the initial sensitivity and specificity were 75 and 70% or the unprocessed and 98 and 82 % for the enhanced images. Conclusions: Often retinal images present with unwanted shadows that may mask pathology to the human grader and the computer–based analysis system. Image processing algorithms are needed to compensate for these artifacts.
This PDF is available to Subscribers Only