Learning Optimal Matched Filters for Retinal Vessel Segmentation with ADA-Boost

Retinopathy of Prematurity (ROP) is an eye disease that affects premature infants. Its signs are tortuosity and dilation of retinal vessels, which are subjectively evaluated by clinicians for the diagnosis and the follow-up of the disease. The availability of algorithms for vascular segmentation would allow vessel geometrical characterization, and hence the quantitative and objective clinical evaluation of the these signs. Unfortunately, algorithms designed for adults’ fundus images do not work well in infants’ fundus images, due to their very low quality. At variance with available methods, we propose a data-driven approach, in which the system learns an array of optimal discriminative convolution kernels, to be employed in a ADA-boost supervised classification. The array is employed as a rotating bank of matched filters, whose response is used by the boosted linear classifier to provide a classification of each image pixel into the two classes of interest (vessel/background). In order to test the generality of the approach, we assessed the performance of the proposed method both on adults’ fundus images using the DRIVE dataset, and also on infants’ images by cross-validation on a dataset of 20 images acquired with a RetCam fundus camera. Average accuracy and Matthews’ correlation coefficient are respectively 0.94 and 0.69 for DRIVE and 0.98 and 0.66 for the Retcam dataset with respect to the manual ground truth references.