Advanced methodologies in multi-shell diffusion-weighted MRI: integrating spherical mean technique with NODDI and improving noise estimation in microstructure fingerprinting

Details

Supervisors

Macq, Benoît

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en mathématiques appliquées, à finalité spécialisée

Abstract

This thesis explores advanced methodologies in Diffusion-Weighted Magnetic Resonance Imaging (DW-MRI) to enhance the analysis and characterization of brain microstructure. In DW-MRI, the processing of DW-MRI signal by complex analytical or numerical models allows the interpretation of the DW-MRI signal into biophysical models that can be related to actual histological compartments in vivo. However, the use of models and the assumptions associated with them introduces biases and errors in the final estimation. This work aims to improve parameter estimations in Neurite Orientation Dispersion and Density Imaging (NODDI) and Microstructure Fingerprinting (MF) by reducing assumptions and biases. The study is divided into two parts: integrating the NODDI model with the Spherical Mean Technique (SMT) for Multiple Sclerosis (MS) analysis and incorporating noise estimation techniques into the MF model. In the first part of this study, we combine NODDI and SMT to enhance microstructural parameter estimation in MS. A key limitation of NODDI is its assumption of fixed parallel diffusivity across voxels. By leveraging SMT for voxel-wise diffusivity estimation, we refine NODDI parameter fitting. Our synthetic experiments demonstrate reduced absolute errors in parameter estimates at higher Signal-to-Noise Ratio (SNR) values (25+). In vivo analysis further shows improved model fit and clear distinctions between white and gray matter regions. Importantly, this method reveals significant differences between the normal-appearing white matter of healthy controls and MS subjects, indicating increased sensitivity to subtle disease-related microstructural changes. The second part of this study aims to enhance the robustness of the MF model by integrating noise estimation techniques. Recognizing that preprocessing alone cannot fully eliminate noise, we introduced a noise parameter to the MF model to prevent diffusion parameters from compensating for noise. Our synthetic experiments show significantly improved fit quality and reduced errors across all SNR levels. In higher SNR settings, our model matches or surpasses ML-enhanced MF models in accurately estimating parameters. Moreover, in vivo data experiments reveal that our method produces parameter estimates more consistent with known brain anatomy and improves the quality of fit, as indicated by higher R2 scores, despite an observed increase in MSE in central brain regions. Overall, this thesis contributes to the field of DW-MRI by proposing innovative approaches to enhance microstructural imaging. The combined NODDI-SMT method offers a promising tool for MS research, while the noise-estimation-enhanced MF model shows potential for broader clinical applications. Future research should focus on optimizing the speed of the fitting process and further improving noise estimation techniques to fully realize the benefits of these methods.