RFS news highlights resources, issues, and news relevant to in-training members of the ACR. If you have a topic idea or would like to contribute to the blog, please email RFS Secretary Patricia Balthazar, MD.
The Way Ahead
Notes from the AI and Medical Imaging workshop
As a trainee, the dynamic perspectives on the application of AI systems throughout the imaging chain have served as fascinating glimpses into the future of the field and the ever-changing landscape of diagnostic and prognostic processes. It’s difficult to miss the idea from conference to conference and publication to publication, but, as an outsider, the direction of the concepts associated with AI in medical imaging may appear fragmented with questions about safety, investigation, implementation, reimbursement, workflow, and regulation. However, in August, multiple professional societies and the National Institutes of Health (namely the National Institute of Biomedical Imaging (NIBIB) and Bioengineering and the NIH Clinical Center), brought together representatives from academia, industry, and government to discuss, debate, and organize thoughts on this topic. As a member of the ACR RFS, I was offered the opportunity to attend by the ACR Data Science Institute™, and the following are some highlights:
The sessions were opened by Krishna Kandarpa, MD, PhD, of the NIBIB with an initial overview of AI in medical imaging offered by ACR DSI Chief Medical Officer Bibb Allen Jr., MD, FACR, ACR DSI Chief Science Officer Keith J. Dreyer, DO, PhD, FACR, Curtis P. Langlotz, MD, PhD, a member of the RSNA Informatics Committee.
According to Dreyer, in a call to the reasonable implementation of AI to enrich the human-machine interface, “If you think about this, we're basically taking AI and human intelligence [and] combining those together, the trick is to define where do we focus attention in narrow AI…for the next 20, 30, 40 years and how do you define those.”
In a panel discussion to close the session, Allen cited the need for transparency in algorithm development, excluding the proverbial black box from the future of AI in diagnostics. This sentiment was furthered by Bradley J. Erickson, MD, PhD, of the Mayo Clinic, in the session centered on foundational research in machine learning and explainable AI. Erickson said, “I think, increasingly, tools will make it so deep learning is not a black box, probably more correctly called an opaque box and we have to learn how to see inside of it.”
Sessions continued with discussion on infrastructure and the processes of machine learning research, as well as the need for quality diverse data sets. In the face of challenges of limited structure in available data for machine learning research and application, Tessa S. Cook, MD, PhD, of the University of Pennsylvania, identified developmental opportunities of “improved methods for automated extraction and [application of] existing ontology, harmonized ontology, to help index and parse and extract meaningful information from the unstructured data with which we work.”
Diving Into Sessions
Ge Wang, PhD, of the Rensselaer Polytechnic Institute, kicked off the first presentation describing methods of applying AI concepts at the point of reconstruction by harnessing the power of raw data and machine learning to outperform current iterative reconstruction processes. Wang also introduced a relatively novel concept he coined “rawdiomics,” or the diagnostic application of machine learning to source raw data.
The latter session on implementation featured DSI leadership combined with representatives from industry and federal entities. Nicholas Petrick, PhD, of the FDA Center for Devices and Radiological Health, discussed a number of available resources and intricacies of the regulatory processes for software, specifically those considered to contain AI functionality, centered on three pillars:
1) Developing clinical association
2) Analytical validation of all software
3) Clinical validation of the tool.
Kevin Lymen, CEO and lead scientist at Enlitic, a clinical AI start-up, outlined ten challenges to clinical AI development from the perspective of industry, citing the clinical nuance of radiological data — a testament to the value of the domain knowledge of the radiologist and clinician.
This final session on Friday was a discussion of the human-machine system with numerous academic and governmental stakeholders contributing. They included representatives from the National Institute of Standards and Technology, National Library of Medicine, National Science Foundation and NIH Clinical Center. Lauded for his leadership and his team’s work in releasing one of the most comprehensive chest radiograph data sets available, Ronald M. Summers, MD, PhD, of the Clinical Center presented on human-computer interaction, citing the challenges the field faces in that “Machine learning offers the hope for a high-performance, generalizable broad based diagnostic cockpit, but against that hope are the continued IT challenges of limited publicly available, robustly labeled, multi-institutional datasets that encompass a wide variety of clinical tasks, and the lack of publicly available robust radiology natural language processing tools makes it hard to do that combination of images and text.”
The workshop concluded with three separate focus group discussions on gaps in foundational machine learning research, machine learning data needs, and AI implementation challenges. While wide-reaching, one of the sentiments that resonated was this sentiment by Langlotz:
“I agree this is an interdisciplinary science… [in] our lab, we view that as kind of the secret sauce…I guess in some ways it's intangible, but some of the things we found work well is that the healthcare folks need to understand how much the computer scientists value great datasets…and the other is I think that you need some people, a small number of people who are catalysts who really have expertise in both areas, and can speak both languages…that helps a lot to bring people together.”
Justin Taylor, MD, is a fourth-year resident in the National Capital Consortium Diagnostic Radiology Program at the Walter Reed National Military Medical Center.
The views expressed herein are those of the author and do not reflect the official policy of the Department of the Navy, Department of Defense, or U.S. government. Additionally, the use of any industrial name(s) does not indicate or imply endorsement by the Department of the Navy, Department of Defense, or U.S. government.