Brain biomarkers from diffusion MRI: correction of motion artifacts, acceleration of microstructure estimation, and disease classification with transformers
(2025)
Files
Thesis_Quentin_Dessain.pdf
Open Access - Adobe PDF
- 32.16 MB
Details
- Authors
- Supervisors
- Macq, Benoît
- Abstract
- Understanding the biological mechanisms underlying brain disorders remains a significant challenge in neuroscience. Traditional diagnostic approaches rely heavily on subjective assessments, leading to variability in diagnosis and treatment outcomes. Diffusion magnetic resonance imaging (dMRI) has emerged as a powerful tool for investigating brain microstructure, offering insights into white matter integrity and connectivity. However, several technical challenges hinder its clinical impact, including motion artifacts, the computational demands of complex tissue modeling and challenges in integrating deep learning with advanced microstructural models for disease diagnosis. This thesis presents methodological innovations to address these challenges across the dMRI pipeline, from acquisition to analysis. The research is structured around three main contributions. First, a motion correction framework tailored for gradient cycling acquisitions in dMRI is introduced, significantly improving data quality in high b-value acquisitions. Second, two deep learning-based approaches are developed to accelerate the estimation of microstructural parameters in Microstructure Fingerprinting. Third, a Swin Transformer-based deep learning model is applied to multi-shell dMRI data for Alzheimer’s disease classification, demonstrating the potential of deep learning in integrating diffusion-based biomarkers for early diagnosis. Together, these contributions advance the practical utility of dMRI by improving data quality, computational efficiency, and clinical applicability.
- Affiliations
Citations
Dessain, Q. (2025). Brain biomarkers from diffusion MRI: correction of motion artifacts, acceleration of microstructure estimation, and disease classification with transformers. https://hdl.handle.net/2078.5/258455