ISMRM25 - Exploring MR Acquisition & Reconstruction: Can We Trust AI as Our Tour Guide?

ISMRM & ISMRT Annual Meeting & Exhibition • 10-15 May 2025 • Honolulu, Hawai'i

Exploring MR Acquisition & Reconstruction: Can We Trust AI as Our Tour Guide?

Weekend Course

ORGANIZERS: HyungJoon Cho, Tolga Cukur, Rita Nunes, Uten Yarach

Sunday, 11 May 2025

311

07:45 - 11:15

Moderators: Onat Dalmaz & Efrat Shimron

Skill Level: Basic to Advanced

Session Number: WE-21

No CME/CE Credit

Session Number: WE-21

Overview
The course will start with basic introductions to the reconstruction of accelerated MRI acquisitions and AI methods. The session will then cover emerging methods for MR acquisition and reconstruction. Finally, trustworthiness issues and the steps required for clinical translation will be discussed.

Target Audience
MR physicists and clinicians who are interested in the development and deployment of novel AI techniques for MR image acquisition and reconstruction.

Educational Objectives
As a result of attending this course, participants should be able to:
• List the basics of MRI acquisition and reconstruction;
• Describe state-of-the-art AI techniques for pulse sequence design and reconstruction;
• Describe explanation and validation approaches to build trust in AI tools; and
• Discuss clinical considerations in deploying AI in radiology practice.

07:45		The Need for Speed: Accelerated MRI Acquisition & Reconstruction View Presentation Video Patricia Johnson Impact: MRI's long scan times motivate acceleration techniques like parallel imaging, iterative regularized reconstruction, compressed sensing, and deep learning. This talk introduces these concepts, and lays the groundwork for deep learning methods in MRI reconstruction.
08:15		Fundamentals of AI Models View Presentation Video DoSik Hwang Impact: A brief introduction to neural networks and deep learning, covering key learning paradigms and core architectures like CNNs, transformers, and diffusion models —setting the stage for upcoming talks on their applications in MR imaging/reconstruction and beyond.
08:45		Emerging AI Methods for Pulse Sequence Design View Presentation Video Matthias Günther Impact: the convergence of deep learning, reinforcement learning, and generative models—supported by robust software tools—paves the way for a new era of MR pulse sequence design.
09:15		Break & Meet the Teachers
09:45		Emerging AI Methods for Image Reconstruction View Presentation Video Philippe Ciuciu Impact: Unrolled deep neural networks produce hallucinations on out-of-distribution data as the training is specific to a given forward operator. PnP models are operator-agnostic and train an equivariant denoiser. Multiscale energy-based models extend PnP.
10:15		Building Trust in Your Novel AI Techniques View Presentation Video Florian Knoll Impact: AI image reconstruction can synthesize realistic looking image features that do not correspond to the actual measurement data. In this lecture, I will discuss strategies how these failure modes can be recognized and avoided so that AI image reconstruction can be trusted in clinical use.
10:45		Translation of AI Technology to Clinical Radiology Dania Daye