Modeling and Correcting Scanner Bias Using Physics- and AI-Driven Method for Reliable Neuroimaging Analysis

Abstract

Diffusion-weighted magnetic resonance imaging (DW-MRI) is a critical modality in neuroimaging for assessing brain structure and connectivity. However, its quantitative reliability is frequently compromised by scanner-induced artifacts and measurement variability. This dissertation addresses these challenges through the integration of physics-based modeling and deep learning to enhance the reliability of neuroimaging analysis. First, we systematically characterize the impact of gradient nonlinearity distortions on diffusion metrics and tractography and propose an efficient two-step approximation correction method. Second, we develop a physics-informed neural network capable of predicting nonlinear gradient distortions from DW-MRI data, eliminating the need for external calibration scans. Third, we demonstrate the impact of gradient nonlinearity and the significance of the correction in clinical dataset. Collectively, this work establishes practical implementation, enabling seamless integration into existing neuroimaging pipelines and advancing the reproducibility of DW-MRI biomarkers.