Section: New Results

Physiological Modeling & Spatio-Temporal Analysis

Participants : Nicolas Honnorat, Sarah Parisot, Stephane Chemouny, Hugues Dufaut, Regis Vaillant, Nikos Paragios.

• Low Gliomas Brain Map: in [33] we introduce a graph-based modeling approach towards spatial position interpretation of low gliomas brain tumors. This was achieved through unsupervised clustering from exemplars, where spatial and geometric proximity of tumors were used to determine the strength connectivity of a graph. Towards automatic estimation of the lowest rank graph that is able to express the observed variation of tumors, an LP problem was solved that determines automatically the number of clusters and their centers while associating the training exemplars with them. Promising results that are well aligned with observations from neuro-sciences demonstrate the potentials of the proposed formulation.

• Coupled Iconic/Geometric Spatio-temporal Segmentation: in [25] we have introduced a combined elongated structures segmentation/tracking approach that was based on a two-layer graphical model. The image layer was exploiting the visual space and was seeking to minimize a data-driven cost while the geometric layers was seeking to establish temporal correspondences of the deforming structure. These two layers were coupled through a common set of variables acting on the deformation of the control points representing the elongated structure. Guide-wire segmentation [24] and tracking in low signal-to-noise ratio interventional images demonstrated the extreme potentials of our approach.