Abstract
Purpose: :
The presentation of a novel fully automated system that separates glaucomatous from healthy cases based on digital fundus images.
Methods: :
A pre-processing step eliminates certain disease independent variations such as illumination inhomogeneities, papilla size differences and vessel structures from the input images. In order to characterize glaucomatous changes, generic feature types (pixel intensities, frequency coefficients, histogram parameters, Gabor textures, spline coefficients) are extracted. In contrast to existing approaches, each feature vector is compressed by Principal Component Analysis. The classification of the transformed features is done by a state-of-the-art nu-Support Vector Machine.For the elaborate experimental evaluation of the proposed system architecture we took a large set of papilla-centered color fundus images of 100 glaucoma patients (FDT test time 67.25 ± 33.4 s) and 100 normals (overall mean age 57.0 ± 10.0 years) from the Erlangen Glaucoma Registry (Kowa non-myd, FOV 22,5°). The gold standard was given by an experienced ophthalmologist based on a complete ophthalmological examination with ophthalmoscopy, visual field, IOP, FDT, and HRT II.
Results: :
Classification of compressed raw pixel intensities gained a success rate of 83% with a specificity of 0.72 and a sensitivity of 0.94 to detect glaucomatous cases. A success rate of 86% was achieved by using spline coefficients with a specificity of 0.78 and a sensitivity of 0.94 to detect glaucoma. The combination of both features slightly increased specificity to 0.82 (sensitivity = 0.92). The kappa statistic of 0.74 states a robust classification scheme.
Conclusions: :
The proposed algorithm achieves a robust and competitive glaucoma detection rate. It is comparable to known methods applied to topographic papilla images and does not depend on segmentation-based measurements. For the first time, automated glaucoma detection is performed on color fundus images. Thus, fundus photography is an appropriate modality for computer-assisted glaucoma screening.
Keywords: image processing • optic disc • detection