AI: Anything Synthetic or Correcting Artifacts
Digital Poster
Acquisition & Reconstruction
Thursday, 14 May 2026
This session presents new usages of AI to improve MRI by creating synthetic images or correcting artifacts.
|
660-04-001.
Physics-Driven Deep Model for Direct T2 Mapping and Distortion Correction in Blip-Reversed Multi-Echo EPI
Impact: The proposed meFD-T2Net enables rapid and accurate correction of susceptibility artifacts in blip-reversed multi-echo EPI, directly producing quantitative T2 maps and distortion-free images within seconds. This capability supports real-time and quantitative imaging, improving feasibility in both research and clinical workflows.
|
||
|
660-04-002.
Deep-learning model for improving B0-induced geometric distortion in clinically acquired Whole-Body Diffusion-Weighted MRI
Impact: Accurate co-registration of whole-body DWI and morphological imaging using deep learning enables cancer characterization through voxel-level multiparametric biomarker measurements. This advancement would refine automatic tumour segmentation to assist in staging and assessment of treatment response in systemic disease.
|
||
|
660-04-003.
Rethinking zero-shot self-supervised learning for MRI reconstruction
Impact: The integration of novel training schemes with lightweight networks enables the feasibility of zero-shot self-supervised MRI reconstruction, offering significant improvements in both performance and time efficiency.
|
||
|
660-04-004.
Model-Based Deep Learning MRI Reconstruction: Adversarial Robustness vs. Network Capacity
Impact: Initial results of this study suggest that larger networks, despite empirically better performance, exhibit greater sensitivity to adversarial perturbations. Therefore, comprehensive evaluation of learned reconstruction networks needs to account for this effect as part of an adequate robustness assessment.
|
||
|
660-04-005.
Motion-Consistent Forward-Distortion Network for Deep Motion-Aware Susceptibility Artifact Correction in EPI
Impact: The proposed mcFD-Net consistently achieves high-fidelity EPI distortion correction across both highly-controlled and realistic levels of motion, demonstrating the effectiveness of the motion-consistent forward-distortion approach. Over two orders of magnitude speedup in correction enables clinically feasible computation times for mcFD-Net.
|
||
|
660-04-006.
Uncertainty-Weighted Consistency Learning for Semi-Supervised Medical Image Segmentation
Impact: We developed a semi-supervised methodology that explores uncertainty information weighting to improve the precision and robustness of medical image segmentation. By utilizing unlabeled data, it reduces reliance on annotated data and offers advantages in scenarios with limited annotations.
|
||
|
660-04-007.
MRA-Based Diagnosis of Intracranial Aneurysm in the era of AI: Consideration for Clinical Practice.
Impact: This study provides insights into the strengths and limitations of
FDA cleared AI applications in MRA based intracranial aneurysm management and
highlights need for ongoing expert oversight and clinical validation before AI
results are fully integrated into patient care.
|
||
|
660-04-008.
3D SNC-PDNet: Adaptive Density Compensation for Robust Non-Cartesian ASL MRI Reconstruction
Impact: The proposed 3D SNC-PDNet introduces 3D non-Cartesian image reconstruction. It improves generalization to unseen contrast, stabilizes ill-conditioned inversions, and provides a memory-efficient, clinically viable framework for accelerated MRI.
|
||
|
660-04-009.
From YouTube to MRI Reconstruction: Overcoming Data Scarcity with Physics-Informed Video Pre-Training
Impact: Leveraging large-scale video data to derive priors for physics-informed
deep learning, we present a strategy to pre-train MRI reconstruction networks
that link natural image statistics with MR physics, enabling generalizable and
data-efficient solutions to the acute challenge of data scarcity.
|
||
|
660-04-010.
From Data Scarcity to Data Synthesis: A Pipeline for Generating Female Pelvic Magnetic Resonance Images
Impact: Generating high-quality synthetic pelvic MRI's with diffusion models enables privacy-preserving data sharing, expands access to diverse and rare disease enhanced training datasets, accelerates development of robust, unbiased AI tools, advancing diagnostic accuracy, research equity, and clinical innovation in women’s imaging.
|
||
|
660-04-011.
Adaptive Correction Diffusion Bridges for Generative MRI Reconstruction in Few Sampling Steps
Impact: By combining physics-informed diffusion priors with efficient adaptively-corrected sampling, ACDB enables high-quality reconstructions at markedly reduced inference cost. These performance and efficiency gains may help realize diffusion-based MRI reconstruction in time-critical clinical settings such as pediatric or motion-prone examinations.
|
||
|
660-04-012.
VHU-Net: Variational Hadamard U-Net for Body MRI Bias Field Correction
Impact: VHU-Net introduces frequency-aware variational learning for bias field correction in body MRI, achieving superior intensity uniformity and tissue contrast across multi-center datasets. It substantially improves downstream segmentation accuracy and enables fast, reproducible quantitative imaging for robust clinical deployment.
|
||
|
660-04-013.
A Cycle-GAN-Based MRI to X-ray and X-ray to MRI Synthesis for the Knee Joint
Impact: Image synthesis
across medical imaging modalities is clinically relevant due to risk factors, healthcare
costs, and resources. Advances in AI networks may provide a solution. MRI to X-ray
to MRI synthesis is an open-ended research question focused in this study.
|
||
|
660-04-014.
Sharp DWI by an optimized combination of complex signal averaging and a complex-domain machine learning denoising
Impact: Sharper diffusion-weighted image was
achieved using an optimized combination of complex signal averaging and machine
learning denoising. The proposed method enables better diagnosis and more
reliable quantitative studies for both clinical and research use.
|
||
|
660-04-015.
Enhancing ULF-dMRI Image Quality Using Tissue-Informed Debiasing Approaches
Impact: Debiasing methods that remove smooth, volume-level biases greatly enhance diffusion contrast and reliability in ultra–low-field MRI, enabling high-quality diffusion tensor imaging on ultra-low field systems and expanding access to advanced neuroimaging.
|