High angular resolution diffusion imaging : processing & visualization

Abstract

Diffusion tensor imaging (DTI) is a recent magnetic resonance imaging (MRI) technique that can map the orientation architecture of neural tissues in a completely non-invasive way by measuring the directional specificity (anisotropy) of the local water diffusion. However, in areas of complex fiber architecture, e.g., crossing fibers, it fails to adequately depict the local diffusion process due to the crude assumption of the underlying diffusion process as Gaussian. To overcome these limitations of DTI, high angular resolution diffusion imaging (HARDI) was introduced enabling a more accurate description of the local diffusion process. However, HARDI has many disadvantages in the acquisition, processing and visualization pipeline that prevents it, for now, from consideration for clinical research practice and application. In particular, for a more accurate description of the underlying diffusion process, HARDI modeling techniques require acquisitions that make the total measurement time 4-5 times longer than for DTI. These acquisitions result in a large amount of data, that is difficult to process and visualize as they require a great deal of computational resources. Additionally, the resultant output of these models is a high-dimensional image that is difficult to interpret for clinicians without technical pre-knowledge. In this thesis some of the drawbacks are addressed and solutions provided: • To find the optimal acquisition schemes for a modern 3T scanner, we perform series of synthetic data experiments as well as corresponding real data MRI scans. We evaluate the angular error in the reconstructed diffusion profiles by analytical decomposition techniques Qball and the DOT. We come to a conclusion that a b-value of 2000s/mm2 and a number of around 70 gradients are sufficient to accurately recover angles of about 60 °. • To simplify the complex HARDI output and reduce the costly post-processing and visualization, we propose a classification scheme based on DTI and HARDI scalar measures. We evaluate the classification power of different measures using a statistical test of receiver operation characteristic (ROC) curves applied to an ex-vivo ground truth crossing phantom. This scheme allows the use of the simple diffusion tensor model where it is justified and uses HARDI only when is necessary. • We use contextual processing of DTI data by considering the influence of the neighboring linear profiles to extrapolate the crossing information. We perform experiments demonstrating that the quality of the extrapolated crossings is similar to the regularized Qball. • We present new interactive visualization tools for HARDI data, where we fuse DTI and HARDI profiles. This visualization reduces the complexity and increases the interactivity with the HARDI data. • At the end, an application of HARDI in neurosurgical research, we present a study of the reconstruction profiles in the area around the subthalamic nucleus that shows the additional value that HARDI acquisitions and modeling can give for deep brain stimulation (DBS).