Curvature detection and segmentation of retinal exudates

In this paper, a segmentation method of the retinal images exudates is proposed. First, pixels that belong to exudates are located using the scale-space extrinsic curvature. These candidate points, are used together with the mean curvature to select possible exudates patches. True exudates are selected using the local maxima blob response through dynamical threshold, which will represent the final segmentation. The proposed scheme is tested with a retinal images public database. The ROC curve is used to validate the final performance, which shows a normalized area under the curve of 96.39%, with a confidence level of 0.8. In that case the sensitivity is 97.07%, the specificity is 99.90% and the accuracy is 99.83%. A final comparison with recent methods is also presented.