Deep Learning Methods to Improve the Utility and Reliability of dMRI Data

Abstract

Diffusion MRI (dMRI) is a powerful neuroimaging technique that provides non-invasive insights into brain tissue microstructures by measuring water diffusion. However, due to various acquisition constraints and the inherent low signal-to-noise ratio, dMRI data often suffer from issues related to both utility and reliability. Specifically, clinical dMRI data acquired with limited angular resolution cannot be fully utilized, reducing its applicability (utility issue). Additionally, dMRI data containing insufficiently corrected artifacts can lead to unreliable measures, compromising the validity of subsequent analyses (reliability issue). These two challenges significantly hinder the productivity of dMRI studies and can result in biased clinical or research outcomes. Addressing both the utility and reliability of dMRI data is therefore critical to improving its overall effectiveness.

In this thesis, we propose two deep learning methods to tackle these issues. The first method, DirGeo-DTI, leverages diffusion gradient information to generate angular resolution-enhanced DTI from a minimal set of diffusion-weighted images, thereby significantly improving the utility of dMRI data. The second method, UdAD-AC, is an unsupervised deep learning framework designed to detect artifacts in dMRI data without requiring annotated training data, thus enhancing the reliability of the data.

Extensive experiments on public datasets demonstrate the effectiveness of both proposed methods, with each achieving superior performance in their respective tasks. Furthermore, evaluations highlight the clinical impact of these methods, demonstrating their feasibility for integration into clinical practice and their potential to significantly improve the quality of dMRI-based research and outcomes.