Section: New Results

Medical imaging

Participants : René Anxionnat, Marie-Odile Berger, Nazim Haouchine, Erwan Kerrien, Pierre-Frédéric Villard, Brigitte Wrobel-Dautcourt, Ahmed Yureidini.

• Vessel reconstruction with implicit surfaces

  This research activity is led in collaboration with Shacra project-team from Inria Lille-Nord Europe and the Department of Interventional Neuroradiology from Nancy University Hospital. It was pursued this year in the context of the SOFA-InterMedS Inria Large-Scale Initiative (http://www.sofa-framework.org/ ).

  Our objective is to offer the interventional radiologists with a patient-based interactive simulator [16] . The medical applications are training to endovascular procedures, planning the intervention, and augmenting the intra-operative images with 3D simulated data. Our contributions address vasculature modeling from patient data, namely 3D rotational angiography (3DRA) volumes. The segmentation should be both user friendly and generate a vascular surface model that is compliant with the computing constraints set in interactive simulation. Within A. Yureidini's PhD thesis, a new model was developed consisting of a tree of local implicit blobby models. The algorithm consists of two steps: first, a vessel tracking step to extract the vessel topology and, second, fitting local surface data points with implicit blobby models at each node point on the vessel centerline.

  An extensive validation of our RANSAC-based vessel tracking algorithm was performed [14] , by comparison with state of the art Multiple Hypothesis Testing  [19] on 10 patient data. Fitting the implicit model to patient data relies on the minimization of a multi-termed energy. A closed form solution was derived, and a blob selection and subdivision heuristic was described to implement an efficient energy minimization algorithm. Both the geometric accuracy and compactness of the resulting vascular models were shown to be excellent [15] .

  Our current goals are: first, to further enhance model compactness by relying on the robustness and versatility of the modeling algorithm and using sparser vascular centerline trees; second, to mathematically ensure the continuity between neighboring local implicit models; and third, to reintroduce the raw image data for a more accurate energy computation, with the aim to design a blobby deformable model.

  This model was implemented in Sofa simulation platform, enabling interactive simulation time and thereby showing an impressive realism during tool navigation. On-going preliminary medical evaluation is being carried on by our fellow interventional radiologist in the framework of intervention planning.

• Designing respiration models for patient based simulators

  The work presented here has been done within a collaboration with Imperial College of London, Bangor University and Inria Aviz team.

  Respiratory models could be a key component in increasing realism in medical simulators. We have previously developed such kind of model. However finding the good parameters to tune the model so that it corresponds to a real patient behavior is not an easy task.

  This year, we have studied methods to automatically tune the elasticity of soft-tissues and the respiratory model parameters based on patient data. The estimation is based on two 3D Computed Tomography scans of the same patient at two different time steps. The parametrization of the model is considered as an inverse problem. Optimization techniques have then been deployed to solve the problem.

  In [13] , we used a random search algorithm to generate a given number of sets of 15 random parameters. The set of parameters that provides the lowest fitness is extracted and corresponds to the solution of the optimization problem.

  In [9] , we have made use of an ad-hoc evolutionary algorithm that is able to explore a search space with 15 dimensions. Our method is fully automatic and auto-adaptive. A compound fitness function has been designed to account for various quantities that have to be minimized. The algorithm efficiency was experimentally analyzed on several real test-cases: i) three patient datasets have been acquired with the “breath hold” protocol, and ii) two datasets corresponds to 4D CT scans. The performance was compared with two traditional methods (downhill simplex and conjugate gradient descent), our random search method and a basic real-valued genetic algorithm. The results showed that our evolutionary scheme provides more significantly stable and accurate results.

• Physics-based augmented reality

  The development of AR systems for use in the medical field faces one major challenge: the correct superposition of pre-operative data onto intraoperative images. This task is especially difficult when laparospic surgery is considered since superposition must be achieved on deformable organs. Most existing AR systems only consider rigid registration between the pre and intraoperative data and the transformation is often computed interactively or from markers attached to the patient's body. In cooperation with the Shacra team, we have introduced an original method to perform augmented or mixed reality on deformable objects. Compared to state-of-the-art techniques, our method is able to track deformations of volumetric objects and not only surfacic objects. A flexible framework that relies on the combination of 3D motion estimation obtained from stereoscopic data and a physics-based deformable model used as a regularization and interpolation step allows us to perform non-rigid and robust registration between the pre and intraoperative images [10] .