Frontiers of AI in Medical Imaging:
End-to-end system design for fast, motion-robust video imaging
Abstract: Magnetic Resonance Imaging (MRI) systems offer rich information about the human body. However, their clinical use is hindered by key challenges, primarily the long scan duration and sensitivity to motion artifacts. Moreover, due to the high complexity of MRI systems, their components are commonly designed separately, which leads to sub-optimal performance. Although machine learning (ML) techniques have recently shown great promise for addressing these challenges, their development is hindered by the scarcity of suitable training data. In this seminar I will introduce new strategies for developing ML-based computational frameworks for fast, motion-robust medical imaging with MRI. First, I will review data-related challenges, and demonstrate that naïve use of open-access medical databases could lead to biased, overly optimistic results. Then, I will present new strategies for rethinking the entire medical imaging pipeline. Specifically, I will introduce two computational frameworks for rapid dynamic (video- capturing) imaging, featuring an end-to-end acquisition-reconstruction design. The first,
BladeNet, combines a motion-informative data sampling technique with a unique ML-based reconstruction network. This framework enables accelerated free-breathing imaging, which is
highly suitable for pediatric patients. The second framework, K-band, addresses challenges in 4D (dynamic-volumetric) imaging by introducing an end-to-end pipeline design, with fast data acquisition and self-supervised reconstruction. This framework enables training self-supervised DL models using only limited-resolution data, which can be acquired rapidly, and offers easy implementation. Finally, I will conclude with an outlook to the future of AI medical imaging, focusing on emerging technologies for low-coast imaging and personalized healthcare.
Bio: Efrat Shimron is a postdoctoral fellow in the Department of Electrical Engineering and Computer Sciences (EECS) at UC Berkeley, working with Prof. Michael (Miki) Lustig. Her research focuses on developing machine learning techniques for medical imaging, focusing on dynamic body imaging. She had previously obtained a PhD from the Technion – Israel Institute of Technology, where she developed Compressed Sensing techniques for rapid MRI. Efrat was selected as a 2023 Rising Star in Electrical Engineering and Computer Sciences (EECS) and 2022 Outstanding Emerging Investigator. She is also the recipient of summa cum laude awards from several international conferences. Additionally, her work on identifying Data Crimes in medical AI algorithms received wide media coverage.