A Multiscale-Multiresolution Method for Automated Retinal Feature Detection

Abstract

Abstract: : Purpose: To develop rapid automated detection and screening for diabetic retinopathy using image analysis methods that apply distinct detection methods to different pixel distributions and extents of retinal features. Methods: Central retinal area color digital images (Topcon) were obtained from subjects with varying severity of diabetic retinopathy and normal subjects and moved via local area network from a remote diabetic clinic to a central server. Images (1024 x 1024 ) were reduced to gray-scale images and cropped to 256 x 256 pixels. Computationally fast and efficient bi-orthogonal wavelets were applied to detect hard and soft exudates. Statistical inference on exudate presence and type was conducted in the wavelet domain. Drusen and micro-aneurysms were detected from raw pixel data using 2-D center-surround filters with a dark pattern inside and a bright pattern outside (drusen) or a bright pattern inside and a dark pattern outside (micro-aneurysms). Blood vessels were detected using Canny's edge detector followed by edge linking. Results: Hard exudates and soft exudates (cotton wool spots) were measured as wavelet coefficients whose values and distribution indicated exudate presence and area. Drusen and micro-aneurysms were measured by filtering and thresholding followed by counting the number of pixels in the output, which indicated the extent of the involved retinal area. Abnormal blood vessels (IRMA and venous beading) were detected by comparison of blood vessel width in diabetic subjects and normals. The patterns of retinal areas with exudates and the pixel distributions of blood vessels with IRMA detected in the manner described are statistically distinct from normal blood vessels of age-matched retinal image controls at a useful level of sensitivity and specificity. Conclusions: The retinal features of diabetic retinopathy can be detected and measured using a novel combination of image analysis methods. These combined methods are computationally rapid, and detect abnormal features and progression of diabetic retinopathy in subjects followed over time. Information obtained from this suite of automated feature detection methods is being applied in an automated statistical learning-based risk assessment of diabetic retinopathy onset and progression.

Keywords: diabetic retinopathy • low vision • imaging/image analysis: non-clinical