Section: New Results

Medical imaging

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

Vessel reconstruction with implicit surfaces

Our 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.

Our objective is the implicit modeling of blood vessels from 3DRA data, with the aim to use these models for real time simulation of interventional procedures. Within A. Yureidini's PhD thesis, a new model was developed consisting of a tree of local implicit blobby models. This model was implemented in Sofa simulation platform, enabling interactive simulation time (60 fps) and thereby showing an impressive realism during tool navigation [20] . We focused this year on the extensive validation of our RANSAC-based vessel tracking algorithm, by comparison with state of the art Multiple Hypothesis Testing  [24] on 10 patient data [18] . Our initial mechanism to fit the implicit model to patient data relies on the minimization of a multi-termed energy. This energy was put under scrutiny, assessing the contribution of each energy term [19] . Our current goal is to reintroduce the raw image data for a more accurate energy computation, with the aim to design a blobby deformable model.

A variational framework for automatic modeling of the vocal tract

Segmenting the vocal tract in MRI is difficult especially because the tongue may move near other edges in the oral cavity, such as the palate or the teeth, which may disturb the segmentation process. The idea explored in our past work was to guide the segmentation with shape priors learnt on a reference speaker within a shape-based variational framework.

Shape priors were incorporated into segmentation via a PCA model with a relatively large number of components to enable the adaptation of the model to strong morphological differences. During this year, this work was continued with the aim to detect tongue contours in physical correspondences, thus allowing us to build a model of the vocal tract. An automatic method for the identification of the end points as well as an improved variational framework to obtain curves in physical correspondences was described in [15] . Second, we extensively assessed the segmentation process. We experimentally showed that the reference model is able to cope with strong morphological differences between speakers with a limited numbers of modes.

Medical simulators based on task analysis

We present here two works done within a collaboration with Imperial College of London.

In order to validate a virtual reality ultrasound-guided targeted liver biopsy procedure simulators previously designed [22] , we have worked on task analysis to deconstruct individual procedural tasks followed by metric definition and critical performance indicator identification. Consultant and trainee scores on the performance metrics were compared. Independent t-tests revealed significant differences between trainees and consultants on 3 performance metrics: targeting, probe usage time and mean needle length in beam. ANOVA reported significant differences across years of experience on seven performance metrics: no-go area touched, targeting, length of session, probe usage time, total needle distance moved, number of skin contacts, total time in no-go area. More experienced participants consistently received better performance scores on all 19 performance metrics [9] .

We used the same task analysis technique to design an inguinal hernia repair simulator [16] . The task analysis allowed to break down the complex operation into sub-tasks and it also provided the foundation for useful and productive discussions between clinical staff and developers. We deployed our system as an e-learning application, allowing surgeons to easily access the application.