Learning a reconnecting regularization term for blood vessel variational segmentation

The segmentation of blood vessels in medical images is a challenging task as they are thin, connected and tortuous. The detection of a connected vascular network is of the utmost importance in clinical applications (e.g. blood flow simulations, vascular network modeling and analysis). Deep learning approaches have been developed to tackle this issue, but they require a large annotated dataset for each new application of interest, which is very challenging to build for vascular networks. In this work, rather than learning the segmentation task, we propose to learn a reconnecting regularization term that learns geometric properties of vascular networks independent of the image modality. Therefore, this term generalizes better than deep learning segmentation models, and can be easily plugged into variational segmentation frameworks to detect vascular networks in different datasets without requiring annotations. We apply this approach on retinal images by training our reconnecting term on the STARE dataset and applying it on the DRIVE dataset. We show that our approach better preserves the connectivity of vascular networks than classic regularization terms in the literature. Finally, we illustrate the generalization power of our reconnecting term by applying it to other types of data.