VisionQuest and its collaborators are engaged in the development of a computer-based system for screening retinal images for various pathologies. The long-term goal of this project is to field a broad-access, reliable, efficient, and cost-effective screening program by implementing a hybrid approach which includes both the application of computer-assisted technology and trained human graders.
Our strategy is to implement a screening system that triages diabetic patients into those with some form of diabetic retinopathy (DR) from those with no signs of DR. The system that will be implemented will aim toward a very high sensitivity and moderately high specificity. Our strategy of near perfect sensitivity and acceptably high specificity ensures that few, if any, diabetes patients with retinopathy are missed by the screening system; yet the high, though not perfect, specificity reduces significantly the number of diabetic patients without retinopathy who are referred to an ophthalmologist or whose images need to be screened by human graders. Our goals for sensitivity and specificity are consistent with, or better than, studies for currently trained “readers” have shown.
The proposed computer-based methodology for DR Screening was based on an approach that closely emulates the human vision system and depends less on lesion-by-lesion segmentation and detection, as do most other approaches that have been reported. Our researchers at VisionQuest explored independent component analysis (ICA) as well as other approaches. One highly innovative approach that was used extensively to produce the results that are reported herein is a methodology called Amplitude Modulation-Frequency Modulation (AM-FM) [1, 2], and was applied by VisionQuest to DR Screening and retinal image analysis by collaborators at the University of New Mexico [3, 4]. Our approach is very different and new, in that a single algorithm has been developed that can be applied to all lesions, any image resolution, any modality (red free, color, fluorescein angiography, etc.), and any retinal disease. A patent is pending for the AM-FM algorithm and screening process.
1. Pattichis, M.S. and Bovik, A.C., “Analyzing image structure by multidimensional frequency modulation,” IEEE Trans. Pattern Anal. Mach. Intell., pp. 753-766, no. 5, May 2007.
2. Pattichis, M.S., Pattichis, C.S., Avraam, M., Bovik, A.C., and Kyriakou, K. “AM-FM Texture Segmentation in Electron Microscopic Muscle Imaging,” IEEE Transactions on Medical Imaging, vol. 19, no. 12, pp. 1253-1258, December 2000.
3. V Murray, M Pattichis, and P Soliz, “New AM-FM Analysis Methods for Retinal Image Characterization,” Asilomar Conference on Signals, Systems, & Computers, Oct, 2008.
4. C Agurto, S Murillo, V Murray, M Pattichis, S Russell, M Abramoff, and P Soliz, “Detection and Phenotyping of Retinal Disease using AM-FM Processing for Feature Extraction,” Asilomar Conference on Signals, Systems, & Computers, Oct, 2008.
