Section: New Results

Image processing on Diffusion Weighted Magnetic Resonance Imaging

Analytic multi-compartment models have gained a tremendous popularity in the recent literature for studying the brain white matter microstructure from diffusion MRI. This class of models require the number of compartments to be known in advance. In the white matter however, several non-collinear bundles of axons, termed fascicles, often coexist in a same voxel. Determining the optimal fascicle configuration is a model selection problem. We have proposed [30] , [33] a novel approach to identify such a configuration from clinical diffusion MRI where only few diffusion images can be acquired and time is of the essence. Starting from a set of fitted models with increasing number of fascicles, we use Akaike information criterion to estimate the probability of each candidate model to be the best Kullback-Leibler model. These probabilities are then used to average the different candidate models and output an MCM with optimal fascicle configuration. This strategy is fast and can be adapted to any multi- compartment model. We illustrate its implementation with the ball-and-stick model and show that we obtain better results on single-shell low angular resolution diffusion MRI, compared to the state-of-the-art automatic relevance detection method, in a shorter processing time.

We have presented [31] a pipeline to reconstruct the corticospinal tract (CST) that connects the spinal cord to the motor cortex. The proposed method combines a new white matter microstructure model coined Diffusion Directions Imaging and a new tractography algorithm based on a particle filter adapted for approximating multi-modal distributions. In this paper, we put the computation time and accuracy of our pipeline to the test in the context of the MICCAI 2014 DTI challenge, which aims to provide fast and accurate reconstructions of the CST for presurgical planning of brain tumor extraction. These two key performance metrics are expected in such a situation where time is of the essence and the quality of the data is dependent on the patient's health condition and ability to cooperate. In no more than 1.5 hours per patient, we successfully provide accurate CSTs of 2 very collaborative patients who underwent a diffusion MRI protocol that included 69 diffusion-sensitizing gradients spread over 4 different shells ranging from b = 200 to b = 3000 $s/m{m}^2$.

Diffusion images are known to be corrupted with a non-central chi (NCC)-distributed noise. There has been a number of proposed image denoising methods that account for this particular noise distribution. However, to the best of our knowledge, no study was performed to assess the influence of the noise model in the context of diffusion model estimation. In particular, multi-compartment models are an appealing class of models to describe the white matter microstructure but require the optimal number of compartments to be known a priori. Its estimation is no easy task since more complex models will always better fit the data, which is known as over-fitting. However, MCM estimation in the literature is performed assuming a Gaussian-distributed noise. We have shown in a preliminary study [32] that using the appropriate NCC distribution for modeling the noise model reduces significantly the over-fitting, which could be helpful for unraveling model selection issues and obtaining better model parameter estimates.