Section: New Results

Handling non-rigid deformations

Participants : Marie-Odile Berger, Jaime Garcia Guevara, Pierre-Frédéric Villard.

Simultaneous pose estimation and augmentation of elastic surface

We have proposed an original method to estimate the pose of a monocular camera while simultaneously modeling and capturing the elastic deformation of the object to be augmented [22]. Our method tackles a challenging problem where ambiguities between rigid motion and non-rigid deformation are present. This issue represents a major barrier for the establishment of an efficient surgical augmented reality where the endoscopic camera moves and organs deform. Using an underlying physical model to estimate the low stressed regions our algorithm separates the rigid body motion from the elastic deformations using polar decomposition of the strain tensor. Following this decomposition, a constrained minimization, that encodes both the optical and the physical constraints, is resolved at each frame. Results on real and simulated data proved the effectiveness of our approach.

Fusing US and CT data

3D ultrasound (3D US) is an ideal imaging modality for hepatic image-guided interventions. Yet, its limited field of view and poor in-depth image quality reduce its usefulness. Within J. Guevara's PhD thesis, we propose to reduce these limitations by augmenting the intraoperative 3D US view with a preoperative image. Our approach is automatic and does not require manual initialization or a tracking device for the 3D US probe. Moreover, by using an underlying biomechanical model, the proposed method handles significant liver deformation, even when it occurs outside the 3D US field of view. The method relies on the segmentation of a vascular tree from the preoperative and intraoperative images, and their transformation into graphs. The preoperative and partial intraoperative graphs are then matched using an algorithm based on a combined Gaussian Process regression approach and biomechanical model. The model is used to robustly select a correct match from several hypotheses generated by the Gaussian Process. Once the two graphs are matched, a deformation of the preoperative liver is driven by the local displacement field computed from the partial graph match.

Individual-specific heart valve modeling

We developed a method to semi-automatically build a mitral valve computational model from micro CT (computed tomography) scans: after manually picking fiducial points on the chordae, the leaflets were segmented and the boundary conditions as well as the loading conditions were automatically defined. Fast Finite Element Method (FEM) simulation was carried out using Simulation Open Framework Architecture (SOFA) to reproduce leaflet closure at peak systole. We developed three metrics to evaluate simulation results. We validated our method on three explanted porcine hearts and showed that our model performs well. We evaluated the sensitivity of our model to changes in various parameters. We also measured the influence of the positions of the chordae tendineae on simulation results.