Section: New Results
Clinical and Biological Validation
Evaluation of registration algorithms on a grid infrastructure
This project is part of a partnership between the Asclepios team, the I3S laboratory and is funded by the AGIR project of the French ACI ``Masses de Données''.
The accuracy of registration algorithms is critical for many clinical procedures but quantifying it is difficult due to the lack of gold standard (ground-truth) in most clinical applications. To tackle this problem, we studied a bronze standard method which relies on a database of in vivo real images representative of the clinical application rather than on simulated or phantom images.
This method considers the ground truth as a hidden variable which is estimated thanks to the redundancy yielded by the computation of the registration of all the possible image pairs of the database with multiple algorithms. This bronze standard estimation maximizes the log-likelihood of the observed transformations.
Accuracy results were obtained on a brain MRI database used for the clinical follow-up of the radiotherapy of brain tumors with four rigid registration methods. Those results showed that the bronze standard method was able to identify non rigidities among the transformations, including image artifacts, high deformations in the tumor area and tilts in the image acquisition. The four studied algorithms exhibited similar sub-voxelic accuracy on this database  ,  .
Because of its computing and data intensive nature, the bronze standard application benefits from a grid implementation which facilitates algorithm and data sharing and provides computing power to speed-up the execution. A deployment of this application on the EGEE and Grid'5000 grids was done using our MOTEUR  ,   workflow manager. MOTEUR allows the execution of an application described as a graph of services in a fully parallel mode  . Specific algorithms were developed to allow data management in this context  .
The execution of the bronze standard application on the EGEE production grid showed that such large-scale and multi-users platforms introduce a high and variable overhead that highly penalizes the application. This overhead was quantified and compared to the one measured on local clusters  which was found to be orders of magnitude lower. However, the throughput of the clusters is also lower than the one of production grids, which makes them saturate as the size of the input dataset increases, thus highlighting the need for a large-scale grid. Thus, we investigated methods to reduce the impact of the overhead on production grids. In particular, job grouping  and job granularity  aim at lowering the number of jobs submitted by the application to reduce the probability for a job to face a high overhead. They are based on a trade-off between the reduction of the number of submitted jobs and the exploitation of the parallelism. Another way to reduce the impact of the grid overhead is to set a timeout value to the jobs and to resubmit them when it expires. We noticed that setting a timeout only speeds up the execution when the tail of the distribution of the grid overhead is sufficiently heavy and when failures impact the execution  . Those strategies were shown to provide a significant speed-up on the execution time and larger scale experiments on the bronze standard application are thus made possible.
Log-Euclidean DTI analysis in clinical studies
The applicative power of the previously proposed Log-Euclidean framework for DTI processing was assessed through different studies lead in collaboration with MD. Denis Ducreux, radiologist at the hospital Kremlin-Bicêtre, Paris. In  , we inspected the usability of DTI and fiber tracking in spinal cord astrocytomas, which are rare neoplasms that can result in alteration of the spinal cord structural integrity (Fig. 25 left). Our objective was to visualize the deformation of the posterior spinal cord lemniscal and corticospinal tracts in 5 patients with low-grade astrocytomas compared with 10 healthy volunteers by using 3D fiber-tracking reconstructions. In  , we showed that DTI, as a pretherapeutic routine investigation in brain tumours, can be helpful as an additional tool to morphological MRI in evaluating the prognosis of patients. Finally, in  , we illustrate the value of diffusion tensor imaging and tractography in the diagnosis and follow-up of central pontine myelinolysis (Fig. 25 right).