Section: New Results
Medical Image Analysis
Estimation of 3D Myocardium Strain from Clinical Cine MRI Using Incompressible Demons
Participants : Tommaso Mansi [ Correspondant ] , Xavier Pennec, Jean-Marc Peyrat, Hervé Delingette, Maxime Sermesant, Nicholas Ayache.
This work has been performed in close collaboration with J. Blanc, MD, and Y. Boudjemline, MD, (AP-HP Necker-Enfants Malades, Paris, France).
The purpose of this research is to include prior knowledge in an efficient non-linear registration algorithm to estimate 3D myocardium strain from standard clinical cardiac cine MRI. The motivation is to have an algorithm that can extract function information for direct clinical applications. The idea consists in constraining diffeomorphic demons, an efficient non-linear registration method previously developed in the team, to provide volume preserving deformations as myocardium is known to be almost incompressible during the cardiac cycle. The incompressibility constraint is expressed in the tangential space of diffeomorphic deformations, which consists in a linear divergence-free constraint on velocities. Elastic-like regularisation is integrated to simulate myocardium visco-elasticity, and a Jacobian-preserving framework is implemented to improve the recovery of incompressible deformations. First results have shown promising results on normal and pathological subjects, when compared with tagged MRI or 2D ultra-sound speckle tracking (see figure 3 ).
|
Regional shape and appearance modeling for deformable model-based image segmentation
Participants : François Chung [ Correspondant ] , Tobias Heimann, Hervé Delingette.
This work is supported by the EU Marie Curie project 3D Anatomical Human (MRTN-CT-2006-035763).
Within the framework of the 3D Anatomical Human project, INRIA is leading the workpackage WT2 which consists in providing algorithms that efficiently extract lower limb structures and motion from static and dynamic medical images [67] , [73] . Our current focus is the appearance description around regions of interest (e.g. lower limb structures, liver, ...) for model-based image segmentation. Instead of relying on Principal Component Analysis such as in Statistical Appearance Models, we propose a method based on Multimodal Prior Appearance Models that does not rely on an accurate pointwise registration [49] . Our method is built upon the Expectation-Maximization algorithm with regularized covariance matrices and includes spatial regularization. The number of appearance regions is determined by a novel model order selection criterion. The prior is described on a reference mesh where each vertex has a probability to belong to several intensity profile classes (see some priors on figure 4 ). This prior's objective is to determine optimal external forces that will guide the deformable model in segmentation approaches.
Analysis and simulation of the heart function from multimodal cardiac images
Participants : Adityo Prakosa [ Correspondant ] , Hervé Delingette, Maxime Sermesant, Tommaso Mansi, Pascal Cathier [ Philips Medical Systems ] , Patrick Etyngier [ Philips Medical Systems ] , Pascal Allain [ Philips Medical Systems ] , Eric Saloux [ CHU Caen ] , Nicholas Ayache, Nicolas Villain [ Philips Medical Systems ] .
This work is done in collaboration with the MEDISYS group of Philips HealthCare, Suresnes, France, and with the University Hospital of Caen, Normandy, France.
In Cardiac Resynchronization Therapy (CRT), used on patients suffering from cardiac motion asynchrony, the selection of patients and the placement of pacemaker electrodes play a crucial role and must be improved since currently 30% of the patients with pacemaker show no benefit from this therapy. We propose to analyze multimodal cardiac images, that are widely available, in order to obtain cardiac mechanical activation times and strains. To this end, we are evaluating the strain estimation based on the incompressible diffeomorphic demons algorithm [59] from volumetric echocardiographics image sequences. Echocardiography myocardium tracking and obtained radial and longitudinal strain are shown in figure 5 .
Automatic detection and segmentation of lesions in medical images
Participants : Ezequiel Geremia [ Correspondant ] , Nicholas Ayache, Olivier Clatz, Antonio Criminisi [ Microsoft Research Cambridge UK ] , Hervé Delingette, Bjoern Menze.
Our method aims to automatically detect and segment various types of lesions in 3D MRI scans. The random decision forest framework provides us with a fast discriminative voxel-wise classifier. In conjuction with a random feature generator, this approach allows a better understanding of relevant features and especially of context based features.
Building generic atlases for radiotherapy planning of the head and neck region
Participants : Liliane Ramus [ Correspondant ] , Grégoire Malandain.
This work is done in collaboration with DOSIsoft S.A., Cachan and Université Catholique de Louvain.
In the context of atlas building for the head and neck region, we proposed a novel method based on kappa statistics to estimate an average segmentation from a database of manual segmentations [61] . We also proposed another method to perform this task from a database of manual segmentations with missing contours. Both methods enable to overcome the over-segmentation obtained with the STAPLE algorithm, as illustrated on Fig. 6 . We also proposed several approaches to design patient-specific atlases, taking into account clinical information such as the localization and stage of the tumor as well as image criteria.
|
Building patient-adapted atlases for radiotherapy planning of the head and neck region
Participants : Olivier Commowick [ Childrens hospital, Boston ] , Grégoire Malandain [ Correspondant ] .
This work is done in collaboration with DOSIsoft S.A., Cachan and Université Catholique de Louvain.
Because of the high inter-subject anatomical variability, using one single generic atlas to segment every subject in the whole population is perhaps unrealistic. We studied here the building of a patient-specific atlas made of pieces of already segmented images. For each region of interest, the most similar image is retrieved among a database, and then all images are fused to generate a virtual image that is further used as atlas to segment the patient at hand [50] .
Spatiotemporal Registration of 4D Time-Series of Cardiac Images
Participants : Jean-Marc Peyrat, Hervé Delingette, Maxime Sermesant, Chenyang Xu [ Siemens SCR ] , Nicholas Ayache.
This work was partially funded by Siemens Corporate Research (NJ, USA) and Microsoft Research (Cambridge, UK).
We propose a novel spatiotemporal registration framework for 4D cardiac CT sequences where the temporal registration aims at mapping corresponding physiological events and where the spatial registration aims at mapping corresponding trajectories of points. By introducing trajectory constraints , the 4D spatial registration problem can be simplified into a 3D multichannel registration problem solved with an extension of the Diffeomorphic Demons [42] to vector-valued images, called Multichannel Diffeomorphic Demons .
A thorough evaluation and comparison with other competing methods was performed on real patient data and synthetic data simulated from a physiologically realistic electromechanical cardiac model [24] . Results showed that the proposed method was the best compromise between registration accuracy, spatial and temporal smoothness of intersequence spatial transformations, and computation times. Moreover, we proposed a new prospective example of application with the spatiotemporal registration of 4D cardiac CT sequences of the same patient before and after radiofrequency ablation (RFA) in case of atrial fibrillation (AF). The intersequence spatial transformations over a cardiac cycle provide a new analysis and quantification of the regression of left ventricular hypertrophy and its impact on the cardiac function. All results were published in a journal article [39] and a PhD thesis [27] .
|
Atlas-based Registration of Brain Images with Tumors
Participants : Hans Lamecker [ Correspondant ] , Marco Lorenzi, Tommaso Mansi, Xavier Pennec.
This project was funded by the European Commission (FP6 - IST-2004-027749: Health-e-Child)
A patient specific simulation of the tumor growth requires the accurate localisation of the tumour and a model of the fibers around it. A standard approach is to register a diffusion tensor atlas to the patient image. Such pediatric DTI atlases could be available through a collaboration with UNC Chapel Hill (USA). One of the difficulties of this approach is to compute (efficiently, accurately and in a robust manner) deformations in the presence of the tumour where there is clearly no correspondence between the atlas and the image. We started developing an efficient atlas-based registration algorithm integrating spatially adaptive confidence weights, first for the special case of binary masks (tumor vs. no tumor). First results demonstrated the feasability (see Fig. 8 ).
|
