Physics-informed AI for virtual magnetic resonance elastography (vMRE)

We will build the AI engine that converts routine diffusion MRI (DTI) into quantitative maps of brain viscoelasticity—shear modulus (G) and damping ratio (ζ)—without the hardware and extra scan time required for conventional MRE. MRE is accurate but rarely available; DTI is routine. Your work bridges that gap.

You will lead the model‑development stream: design and train physics‑informed neural networks/neural operators that combine data‑driven learning with the equations of motion (momentum balance, constitutive constraints), incorporate voxel‑level uncertainty weighting, and learn robustly from multi‑site DTI via lightweight harmonisation. You will experiment with architectures (e.g., time‑slice U‑Nets, Conv‑LSTMs, neural operators), enforce physical plausibility, and optimise for clinic‑speed inference.

Training resources include a 10,000‑case simulation library, an augmented digital‑phantom set with realistic noise and heterogeneity, and an independent human cohort with paired DTI+MRE (n≈100) for calibration and blinded testing. You will benchmark with pre‑specified primary metrics (ICC, Bland–Altman limits, bias), run ablations to understand failure modes, and package the final model as an ONNX/Docker service for research deployment.

The position is computational and collaborative. You will work in Python, use GPU clusters (NeSI/ABI), follow good ML practice (versioned data/code, containers, reproducible training), and engage with imaging scientists and clinicians. Expected outputs include first‑author papers, open‑source code, and a validated vMRE engine ready for clinical‑pilot studies.

We welcome applicants with a first‑class Honours/MSc (or equivalent) in Biomedical/EE/Software Eng, Computer Science, or Applied Mathematics.

Required: strong ML and Python; desirable: MRI/DTI, signal processing, or PDE‑constrained learning.

Marsden

Page expires: 17 May 2026