Probing Brain Tissue Microstructure with Magnetic Resonance Imaging through Simulation and Bayesian Learning of Signal Dynamics

Event Date:
2026-07-28T01:00:00
2026-07-28T13:00:00
Event Location:
DMCBH 3502 (Djavad Mowafaghian Centre for Brain Health - 2215 Wesbrook Mall)
Speaker:
Jonathan Doucette, Departmenal Defense
Related Upcoming Events:
Grad Theses
Intended Audience:
Everyone
Local Contact:

gradcoord@phas.ubc.ca

All are welcome to this defense!

Event Information:

Abstract:

Magnetic resonance imaging (MRI) resolves brain structures at the millimetre scale, up to three orders of magnitude coarser than the blood vessels, perivascular spaces, and myelinated axons of white matter. This microstructure cannot, therefore, be imaged directly; in this thesis, I combine physics-based simulation with statistical and deep-learning inference to develop methods for inferring these structures' properties from their imprint on the magnetic resonance signal.

First, I developed an operator-splitting solver for the Bloch-Torrey equation that is one to two orders of magnitude faster than direct methods, enabling simulation of spin-echo and gradient-echo dynamic susceptibility contrast in three-dimensional white-matter voxels containing blood vessels and perivascular spaces. These simulations suggested that roughly half of the white-matter blood volume resides in anisotropic vessels aligned with the fibre tracts, and that perivascular diffusion enhances the signal's orientation dependence. Next, I built a finite-element framework for simulating multi-spin-echo signals from myelinated axons. These simulated signals were used to train a conditional variational autoencoder to estimate axon-scale tissue parameters, targeting myelin water fraction, g-ratio, and relaxation times.

Turning to whole-brain analysis, I created DECAES, an open-source regularized nonnegative least squares toolbox that reduces myelin water mapping time from hours to minutes to seconds. Because MRI magnitude data is Rician-distributed and quantized, I developed fast, numerically stable routines for the (quantized-)Rician log-likelihood and its derivatives, improving signal modelling at low signal-to-noise ratio and enabling differentiation through likelihood-based inference.

Finally, these tools enabled Bayesian inference of bi- and multi-exponential Rician signal models. For the former, I developed a semi-supervised approach that embeds a Metropolis-Hastings step to incorporate real MRI signals into training alongside simulated signals, improving real-data inference performance. For the latter, I trained a conditional normalizing flow over the probability simplex to precondition scalable Markov chain Monte Carlo sampling. On synthetic and in vivo data, this framework reduced myelin water fraction error by roughly 30-40% relative to DECAES, produced well-calibrated credible intervals, improved sampling efficiency by one to three orders of magnitude, and enabled further applications such as Bayesian flip-angle estimation and inter-voxel spatial regularization.

Add to Calendar 2026-07-28T01:00:00 2026-07-28T13:00:00 Probing Brain Tissue Microstructure with Magnetic Resonance Imaging through Simulation and Bayesian Learning of Signal Dynamics Event Information: Abstract: Magnetic resonance imaging (MRI) resolves brain structures at the millimetre scale, up to three orders of magnitude coarser than the blood vessels, perivascular spaces, and myelinated axons of white matter. This microstructure cannot, therefore, be imaged directly; in this thesis, I combine physics-based simulation with statistical and deep-learning inference to develop methods for inferring these structures' properties from their imprint on the magnetic resonance signal. First, I developed an operator-splitting solver for the Bloch-Torrey equation that is one to two orders of magnitude faster than direct methods, enabling simulation of spin-echo and gradient-echo dynamic susceptibility contrast in three-dimensional white-matter voxels containing blood vessels and perivascular spaces. These simulations suggested that roughly half of the white-matter blood volume resides in anisotropic vessels aligned with the fibre tracts, and that perivascular diffusion enhances the signal's orientation dependence. Next, I built a finite-element framework for simulating multi-spin-echo signals from myelinated axons. These simulated signals were used to train a conditional variational autoencoder to estimate axon-scale tissue parameters, targeting myelin water fraction, g-ratio, and relaxation times. Turning to whole-brain analysis, I created DECAES, an open-source regularized nonnegative least squares toolbox that reduces myelin water mapping time from hours to minutes to seconds. Because MRI magnitude data is Rician-distributed and quantized, I developed fast, numerically stable routines for the (quantized-)Rician log-likelihood and its derivatives, improving signal modelling at low signal-to-noise ratio and enabling differentiation through likelihood-based inference. Finally, these tools enabled Bayesian inference of bi- and multi-exponential Rician signal models. For the former, I developed a semi-supervised approach that embeds a Metropolis-Hastings step to incorporate real MRI signals into training alongside simulated signals, improving real-data inference performance. For the latter, I trained a conditional normalizing flow over the probability simplex to precondition scalable Markov chain Monte Carlo sampling. On synthetic and in vivo data, this framework reduced myelin water fraction error by roughly 30-40% relative to DECAES, produced well-calibrated credible intervals, improved sampling efficiency by one to three orders of magnitude, and enabled further applications such as Bayesian flip-angle estimation and inter-voxel spatial regularization. Event Location: DMCBH 3502 (Djavad Mowafaghian Centre for Brain Health - 2215 Wesbrook Mall)