Curvilinear Structure Tracking Based on Dynamic Curvature-penalized Geodesics

Abstract

Geodesic models are considered as a fundamental and powerful tool in the applications of curvilinear structure extraction, where the target structures are usually modeled as geodesic paths connecting prescribed points. Despite great advances in geodesic models, it still remains an unsolved problem of detecting weak curvilinear structures from complicated scenarios. In this paper, a dynamic high-order geodesic model for curvilinear structure extraction is introduced to alleviate the shortcuts or short branches combination problems suffered in the classical geodesic approaches. For that purpose, we take into account the nonlocal pattern of curvilinear structures and the local curvature of geodesic paths for the construction of geodesic metrics. Accordingly, the proposed model is able to blend the benefits from the on-the-fly nonlocal smoothness property, curvature regularization and appearance coherence penalization. The nonlocal smoothness property carried out via a local bending operator is constructed to provide a quantitative measure of geodesic advancing directions, meanwhile the coherence penalization is established to guarantee the consistency of the local appearance features extracted via a vessel detector. The experiment results on synthetic and real images illustrate that the proposed method obtains outperformance when compared to the classical geodesic-based tracing algorithms.