Discrete Methods for Guide-wire Segmentation and Tracking

In this report, we propose a complete pipeline for guide-wire segmentation and tracking in the context of cardiac angioplasty. Segmentation is addressed through a bottom up approach which first detects interest points using appropriate features. These points are then clustered into local segments using geometric and proximity constraints. Ordering of these segments through a local permutation model is finally carried out towards complete delineation of the guide-wire. Tracking is addressed using a discrete iconic-geometric approach. Control points are used to represent the guide-wire. Their displacements in time aims at optimizing an image-driven likelihood (on the same space used for segmentation), preserving the geometric characteristics of the curve while establishing correspondences for meaningful landmark points automatically extracted along the guide-wire. The tracking and the landmark correspondence problem form a two layer inter-connected graphical model, that is solved using linear programming towards simultaneous recovery of the two set of variables. Large scale clinical validation and comparisons with state of the art methods demonstrate the potentials of our method.