The segmentation of 2D and 3D vascular networks from medical images has been a florishing field in the past 20 years. The knowledge of the morphology, geometry and topology of vascular networks is an important help for clinicians for diagnostics, chirurgical planning and patient monitoring. The vascular network variability for one patient at different times or from one patient to another calls for robust segmentation algorithms.

After an analysis of the state of the art in vascular segmentation, we identified that a number of algorithms included a vesselness filter that is able to increase the contrast of tubular structures. These filters greatly impact the quality of the final segmentation, however their performance are rarely evaluated outside of a segmentation pipeline. We propose a comparison of 7 vesselness filters in a common comparison framework and implementation. We also propose a granular analysis of the filters performances at the organ scale, the vessels neighbourhood for different vessels size and vessels bifurcations.

One of the main objectives of this work is the extention by the community. That is why, we focussed on the reproducibility of the experimentations, as well as giving access to the code and choosing public datasets.

A second objectif is the combined usage of vesselness filters and deep learning for the extraction of vessels centerlines in a context where only bifurcations are annotated.