1

A diffusion transition process for accelerated MRI reconstruction that is anchored on low-frequency k-space data.

A diffusion model for MRI reconstruction whose forward process progressively removes k-space frequencies while adding Gaussian noise, mimicking undersampled acquisitions.

Super-Resolution Diffusion Bridges (SRDB) learn a direct mapping from low- to high-resolution MR images instead of starting from Gaussian noise, improving structural consistency.

Fourier-constrained diffusion bridges for accelerated MRI reconstruction. Oral presentation — Summa Cum Laude Merit Award.

FDB learns a multi-step mapping from undersampled to fully-sampled MRI data via noise addition and random frequency removal, outperforming prior diffusion priors on brain MRI.

Diff-Med-Synth combines denoising diffusion models with GANs to generate realistic and diverse medical images from noise with fast sampling.

ResDiff combines a resolution-decreasing operator with diffusion noise scheduling, achieving competitive MRI reconstructions in far less sampling time.

A personalized federated learning approach for multi-contrast MRI translation across institutions with heterogeneous data.

A personalized federated learning approach that trains one unified MRI contrast synthesis model across institutions while adapting to each site and task.

pFLSynth uses site- and task-specific latents, personalization blocks and partial model aggregation to train one synthesis model that generalizes across sites and contrasts.