Differentiable Physics for MRI

Why it matters

MRI simulation predicts the signal produced by an experiment. Inverse problems ask the reverse question: which sequence produces a desired response, or which image explains the measured data? Both require repeatedly evaluating a physical model and understanding how its output changes with its inputs. Fast, accurate simulations and automatic differentiation make these optimization problems more practical.

Contribution

I work on making GPU-accelerated MRI simulations differentiable and accurate enough to use inside optimization loops, connecting sequence design and image reconstruction through a common physics-based approach.

RF pulse and sequence design

With Kareem Fareed and collaborators, we use compiler-level reverse-mode automatic differentiation to optimize RF pulses through GPU-accelerated MRI simulations. Our ISMRM abstract presents this approach.

My JuliaCon talk, How I Drew the Julia Logo Using Spins in an MRI Machine, demonstrates it on a scanner: an optimized RF pulse excites a Julia-logo pattern in a water-bottle phantom, using KomaMRI for simulation and Pulseq for acquisition.

Accurate simulations for optimization

I developed the theory and GPU kernels behind our Magnus-based Bloch simulations. This work improves the speed and accuracy of RF excitation simulations and supports their use in inverse design, where numerical errors in the forward model can lead an optimizer to an incorrect solution.