Purchase this article with an account.
G. N. Holland, N. Manukian, P. S. Kalyani, D. C. Yoon, N. Keorochana, J. D. Keenan, S. Ausayakhun, C. Jirawison, S. Ausayakhun, T. P. Margolis; Development of Pattern Recognition Software for Automated Computer Detection of AIDS-Related Cytomegalovirus Retinitis on Digital Fundus Photographs. Invest. Ophthalmol. Vis. Sci. 2010;51(13):4852.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To explore the feasibility of developing pattern-recognition software for automated identification of AIDS-related cytomegalovirus (CMV) retinitis via telemedicine. In the first phase of development, we established and tested algorithms to distinguish between normal fundus and CMV-infected regions on digital fundus photographs (FP).
Experienced ophthalmologists identified "satellite" border regions typical of CMV retinitis on 3 representative FP. We developed feature-detection algorithms to identify the distinctive characteristics of these regions, which consist of small, roughly circular "dots" with a range of sizes and colors, and to detect clusters of dots with interlesional distances typical of those in the sample regions. A matched filter method was tuned to detect these features and to distinguish them from background. The algorithms were then used on a different set of 15 FP having a variety of CMV retinitis lesions, to determine the test sensitivity (SE), specificity (SP), positive predictive value (PPV), and negative predictive value (NPV) for distinguishing between normal fundus and involved retina. Specifically, two rectangular windows (one encompassing a CMV retinitis lesion; one without involved retina) were chosen on each FP to be assessed by the software for the presence of CMV-infected tissue within the windows. In a second study, we placed the rectangle in the "burned-out" central area of large lesions.
The algorithm correctly detected 13 of 15 windows with CMV retinitis lesions (SE=86.7%) and incorrectly detected disease in 2 of 15 windows without CMV retinitis (SP=86.7%). The PPV and NPV were both 86.7%. In the second study, the algorithms identified inactive scar as being diseased tissue, despite the absence of typical satellites, in 3 of 5 windows (SE=60%).
The matched filter technique appears to be an appropriate method for identification of CMV retinitis lesions. Furthermore, results suggest that it should be possible to refine the software to discriminate between active disease and inactive scars. Automated detection via telemedicine will be useful for monitoring people with AIDS for development of new CMV retinitis lesions and for reactivation of treated lesions in regions of the world where access to ophthalmic care is limited.
This PDF is available to Subscribers Only