Retinal fundus image constrast normalization using mixture of Gaussians

We present a fast and robust method to correct contrast variation in retinal fundus imagery. The technique uses a mixture of Gaussians to model the bias of the intensity variation. Typically a three or four component mixture is sufficient to characterize the principal variation due to the spherical geometry of the retina, the high-contrast reflection off the optic nerve and the darker macula. We compare the results with a non-parametric, filtering approach on a standard diabetic retinopathy database of 89 images. Our results indicate that a parametric approach using mixture Gaussian is better at contrast stretching in lesion regions making is an effective pre-processing step for manual and computer aided diagnostic techniques.