Scale-space curvature detection of retinal exudates with a dynamic threshold

In this work a robust and fully automatic method for the detection of hard exudates and soft exudates in ocular fundus images is proposed. Ocular fundus images are represented in scale-space and the curvature is computed in each scale. Curvature extremes locations are determined and then selected with dynamic thresholding in each scale. Those extremes are tracked along the scale-space and their initial locations in the initial image are computed. These locations are likely to be inward the retinal exudates. This algorithm provides a reliable detection and marking of both types of exudates. The final performance is represented by the ROC curve, which shows a normalized area under the curve of 98.43 % for both exudates with a confidence level of 0.8. In that case the sensitivity is, 98.51%, the specificity is 99.6 % and the accuracy is 99.89%. A final comparison with recent methods is also performed.