Impact: Higher resolution in MRE is essential for identifying heterogenous substances and accurately quantifying stiffer tissues, but is typically infeasible by the prolonged scan times. This work enabled fast MRE by reconstructing images from highly undersampled data using a deep-learning-based approach.

Impact: We propose a distribution-adaptive deep learning framework for accelerated low-field MRI. The model needs only high-field data to pre-train a robust diffusion prior and is then lightly adapted in situ, enabling high-fidelity low-field reconstruction without paired high-/low-field datasets.

Impact: 
Multiscale Energy-based models (EBMs) are diffusion models whose score is a potential gradient. They offer: explicit priors for inverse problems, convergent algorithms, and conservative scores. The proposed distillation method enables training of larger EBMs on bigger datasets, offering improved performance.

Impact: PCDM enables high-fidelity multi-contrast MRI reconstruction from highly undersampled data without ground-truth training, effectively mitigating domain shift and offering strong potential for clinically feasible rapid quantitative imaging.

Impact: Proposed CM-RED enables high-quality MR reconstruction with substantially reduced computational cost, making generative AI priors practical for clinical and research applications. By demonstrating that CMs can serve as powerful learned priors, this work enables fast and stable physics-guided generative reconstruction.

Impact: We employ high-quality T1WI as reliable anatomical guides to accurately reconstruct T2 images without needing paired T1-T2 datasets or high-quality T2 supervision, potentially mitigating clinical burden of prolonged T2 scan durations and reducing reliance on scarce high-quality T2WI resources.

Impact: A novel method for improving the quality of complex-valued images in diffusion model reconstruction has been proposed. Phase-varied images can be reconstructed using a real-value trained diffusion model, enabling high-quality images regardless of phase changes.

Impact: The proposed diffusion-based reconstruction method, DDfire, enables high-fidelity MRI reconstruction by whitening denoiser input error via colored “renoising” within a diffusion sampler. For brain MRI, it improves perceptual and distortion image quality metrics at R=4 and R=8 over state-of-the-art methods.

Impact: We integrate implicit neural representations, optical flow, and RED-diffusion priors to enhance dynamic MRI reconstruction. This method learns the distribution of real images, preserves structural details, shows strong noise robustness, reduces artifacts, and improves overall temporal consistency and reconstruction quality.

Impact: By integrating diffusion priors with data-informed convex optimization tools, OptDiff (our method) delivers high-quality MRI reconstructions at a fraction of the computational cost.

Impact: Recent advancements in hardware increase interest in low-SNR MRI systems. Traditional techniques designed to accelerate MR imaging acquisitions may be inapplicable with noisy input. Here we propose a deep-learning, diffusion-bridge–based framework that jointly reconstructs and denoises undersampled, noisy k-space data.

Impact: We provide evidence that an accelerated IR-UTE MRI sequence can be used to detect a tibial plateau fracture.

Impact: Enabling quantitative pharmacokinetic analysis of low tRes clinical breast DCE-MRI data through Gen-AI enhancement of tRes with high fidelity would astronomically increase the data volume for quantitative breast DCE-MRI research and facilitate its translation into clinical practice for precision medicine.

Impact: This work demonstrates that synthetically supervised deep learning can improve perceptual quality and reduces artefacts of dynamic speech MRI. This mitigates the spatiotemporal resolution trade-off, enabling clearer visualisation of rapid articulatory motion for clinical assessment of speech.

Impact: Weakly supervised training releases the dependence on fully sampled data. Our approach ensures consistent dual-contrast reconstruction and phase-sensitive UNI combination for highly accelerated non-Cartesian MP2RAGE data, enhancing the practicality of using fast 7T MP2RAGE imaging.

Impact: Patch-based diffusion inverse solver (PaDIS) improves T2-weighted prostate MRI image quality and artifacts suppression at both 3T and 0.55T. This can potentially enable higher-quality, faster prostate MRI in routine clinical workflows, and accelerate low-field prostate MRI adoption.