Common vessel centerline extraction methods rely on the computation of a measure providing the likeness of the local appearance of the data to a curvilinear tube-like structure. The most popular techniques rely on empirically designed (hand crafted) measurements as the widely used Hessian vesselness, the recent oriented flux tubeness or filters (e.g. the Gaussian matched filter) that are developed to respond to local features, without exploiting any context information nor the rich structural information embedded in the data. At variance with the previously proposed methods, we propose a completely data-driven approach for learning a vesselness measure from expert-annotated dataset. For each data point (voxel or pixel), we extract the intensity values in a neighborhood region, and estimate the discriminative convolutional kernel yielding a positive response for vessel data and negative response for non-vessel data. The process is iterated within a boosting framework, providing a set of linear filters, whose combined response is the learned vesselness measure. We show the results of the general-use proposed method on the DRIVE retinal images dataset, comparing its performance against the hessian-based vesselness, oriented flux antisymmetry tubeness, and vesselness learned with a probabilistic boosting tree or with a regression tree. We demonstrate the superiority of our approach that yields a vessel detection accuracy of 0.95, with respect to 0.92 (hessian), 0.90 (oriented flux) and 0.85 (boosting tree).