Motion Estimation and Inverse Problems in Tagged MRI
Physics-constrained deep learning methods for motion estimation, tracking, and blind inverse problems in tagged MRI, with applications to tongue and cardiac imaging.
This is the central thread of my PhD research, developing deep learning methods for motion estimation and inverse problems in tagged MRI.
DRIMET (MIDL 2023, Oral)
DRIMET is a deep registration-based method for 3D incompressible motion estimation in tagged MRI. It introduces sinusoidal phase transformation and an incompressibility constraint to enable biologically plausible deformation estimation. Validated on tongue tagged MRI datasets of healthy individuals and glossectomy patients, demonstrating utility in assessing speech dynamics post-oral cancer treatments.
MomentaMorph (MICCAI Workshop 2023)
MomentaMorph extends DRIMET for Lagrangian motion estimation by leveraging Lie algebra and Lie group principles to use temporal information, improving tracking amid large motions and repetitive patterns in tagged MRI.
BITE (IPMI 2025)
Brightness-Invariant Tracking Estimation addresses the fundamental challenge of brightness changes in tagged MRI sequences, enabling more robust motion tracking over longer time horizons.
Physics-Constrained Generative Model (CVPR 2026)
Our latest work develops a novel physics-constrained generative model (DDPM) to solve the blind inverse problem of disentangling tagged MRI into cine MRI and tag patterns while simultaneously estimating motion, achieving unprecedented accuracy in motion tracking.
This research has been supported by a JHU Discovery Grant ($150K) and has led to a patent application.