Back to search

Cornell University

Advanced Methods For Magnetic Resonance Image Reconstruction, Motion Artifact Correction And Quantitative Susceptibility Mapping

Abstract

dc:description.abstract

Advances in quantitative MRI increasingly rely on methods that can overcome motion, extract multiple tissue parameters efficiently, and provide clinically meaningful biomarkers. Across the brain and liver, motion compensation, multi-parametric mapping, and susceptibility-based tissue characterization remain key challenges. In this work, we present three complementary developments that address these challenges: deep-learning–based motion correction for mGRE imaging, free-breathing whole-liver quantitative mapping using implicit neural representations, and susceptibility-based fibrosis assessment validated against histology. We first propose a temporal fusion network (tUnet) for motion correction of multi-echo gradient-echo (mGRE) images and further refine it using autofocus-derived, subject-specific motion trajectories to create a fine-tuned version (tftUnet). Evaluated in the context of quantitative susceptibility mapping (QSM) for Parkinson’s disease (PD), the method significantly improves image quality. In simulated brain data (n = 15), tUnet boosted SSIM, PSNR, and RMSE (all p < 0.05). In nine PD patients with real motion artifacts, radiologist scoring increased from 1.11 ± 0.33 (uncorrected) to 2.56 ± 1.01 (tUnet) and 3.33 ± 1.22 (tftUnet), demonstrating substantial suppression of motion artifacts in clinical QSM. Motivated by the broader need for motion-robust quantitative imaging, we also developed a free-breathing 3D whole-liver multi-parametric mapping technique capable of estimating water T1, water T2, fat fraction (FF), and R2*. Using a multi-echo 3D stack-of-spirals acquisition with inversion-recovery and T2-preparation, we designed an implicit neural representation–based fingerprinting method (FINR) that jointly reconstructs static images, motion fields, water–fat separation, and parametric maps. In ten healthy subjects, FINR produced sharp, motion-free images with minimal bias compared with conventional breath-held imaging, with training requiring ~3 hours per subject and inference under 1 minute. Finally, to enable non-invasive assessment of chronic liver disease, we investigated whether mGRE-derived susceptibility sources can quantify fibrosis in ex-vivo liver explants. After fat–water separation, mGRE data were processed to obtain R2QSM and R2’QSM (R2’ = R2 – R2), isolating diamagnetic (negative) susceptibility related to fibrosis. In 20 fixed liver sections with histology, negative susceptibility differentiated mild (F0–F1) from moderate–advanced fibrosis (F2–F3; p = 0.0025), distinguished F2–F3 from cirrhosis (F4; p = 0.021), and separated early (F0–F2) from advanced disease (F3–F4) with 90% sensitivity, 90% specificity, and an AUC of 0.88. It outperformed other MRI parameters and showed positive correlations with fibrosis stage (r = 0.60 for R2*QSM; r = 0.58 for R2’QSM).

Degree

thesis:*
Name thesis:degree_name
Ph. D., Applied Physics
Level thesis:degree_level
Doctor of Philosophy
Discipline thesis:degree_discipline
Applied Physics
Grantor
Cornell University
Year dc:date.issued
2025

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Li, Chao
Chair dc:contributor.chair
  • Zipfel, Warren
Committee members dc:contributor.committeemember
  • Wang, Yi
  • Xu, Chunhui

Rights

Language dc:language.iso
en

Identifiers

dc:identifier.*
Dc Identifier Bibid
17423377
Dc Identifier Other
ProQuest Submission ID: 15370
ProQuest Publication ID: 32286965

Chain of custody

source
Harvested from
Cornell University
Base URL
ecommons.cornell.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Li, Chao. Advanced Methods For Magnetic Resonance Image Reconstruction, Motion Artifact Correction And Quantitative Susceptibility Mapping. Doctor of Philosophy thesis, Cornell University, 2025. https://hdl.handle.net/1813/121016