Feature | Artificial Intelligence | July 12, 2023 | By Christine Book

AI Essentials in Radiology: Experts Weigh In

Leading radiologists offer insight into AI governance and clinical applications

Getty Images

As the radiology community continues to learn and leverage the latest advancements in the use of artificial intelligence (AI), it was fitting that an esteemed panel of leading experts in this area presented an important and comprehensive session on the final day of the 108th Radiological Society of North America (RSNA) Scientific Assembly and Annual Meeting.

One of the hottest topics during the conference, combined with the line-up of Katherine P. Andriole, PhD, Linda Moy, MD, Dania Daye, MD, PhD, and Walter F. Wiggins, MD, PhD, generated strong interest and participation. An overview of basic terms, level-setting of key challenges, cautionary notes on clinical applications and forecasting future applications were well-delivered and well-received.

The presence of these individuals and this session, “Back to the Basics: What Do Rads Need to Know About Radiology AI?” was a logical follow up to the AI Fireside Chat presented during RSNA 2021, and a reflection of the fast pace of growth in AI software solutions and clinical applications in the market today.

ITN’s editorial team is sharing a multi-part overview of information presented. This first segment includes excerpts from Drs. Daye and Wiggins. Daye, with Wiggins and other authors, recently published a paper in RSNA’s journal, Radiology, which she reviewed. In a forthcoming second segment, ITN will publish instructional and insightful session segments from Drs. Andriole and Moy.

Setting the Stage for Greater Understanding

Prior to the session, ITN spoke with moderator and panelist Daye, an interventional radiologist at Massachusetts General Hospital and faculty member at Harvard Medical School, and at the MGH/HST Martinos Center for Biomedical Imaging.

Daye shared the following insights with ITN on the session’s goals.

“I think for the last couple of years we have had such an explosion of AI applications we are seeing in the market, and if you look around, there are so many people using AI in their clinical practice. Yet, we still see so many radiologists out there who are not really familiar with the basics of AI. We’ve seen a huge interest in becoming more familiar and really learning about the basics of implementation. So, the purpose of this session is trying to get the very basics to the people who are just starting to jump into this area and learn a little bit more about data science, implementation and governance, and really take a look at what is out there and how it can be used to provide an infrastructure to get people started. This is surely the beginning of a very long journey, but we hope to get people to dip their toes in the water, so to speak, and help get them started.”

Addressing AI Governance: Who Decides and How?

In introducing the current focal points for the panel discussion on the use of AI in radiology, Daye’s participation in the panel focused on an RSNA journal Radiology paper (published online August 2, and in print December 2022), titled, “Implementation of Clinical Artificial Intelligence in Radiology: Who Decides and How?” for which she was lead author. Daye was joined by Wiggins and other leading radiologists, including James A. Brink, MD, a 2023 RSNA Gold Medalist, in developing the paper.

Focusing on key takeaways from the paper, Daye emphasized that clinical imaging AI programs have four key requirements for successful implementation, which are: data access and security; cross-platform and cross-domain integration; clinical translation and delivery; and leadership supporting innovation. She detailed the needs in this way:

•Data access: Construct an environment with data security protecting patient information, recognizing complete de-identification is difficult.

•Data science: Equip clinical researchers with the new training and tools of data science to harness this data.

•Cross modality: Combine access to EMR, medical imaging, genomics and physiological monitoring data in one location.

•Delivery: Bring new discoveries into clinical care through context integrated clinical decision support platforms via process that aligns across healthcare enterprise.

Daye reinforced the key takeaways from the paper in offering the following: “Successful clinical implementation of artificial intelligence is facilitated by establishing robust organizational structures to ensure appropriate oversight of algorithm implementation, maintenance and monitoring.” The authors also wrote: “As the role of artificial intelligence in clinical practice evolves, governance structures oversee the implementation, maintenance and monitoring of clinical AI algorithms to enhance quality, manage resources and ensure patient safety.”

Highlighting the value of the article, Daye explained that she and the co-authors established a road map for governance, answering four key questions: Who decides which tools to implement? What factors should be considered when assessing an application for implementation? How should applications be implemented in clinical practice? and How should tools be monitored and maintained after clinical implementation? Devising flexible governance structures that can quickly adapt to a changing environment will be essential to ensure quality patient care and practice improvement objectives, she noted.

Daye further emphasized that oversight of AI in medical imaging should consider stakeholders across multiple disciplines who utilize radiology services. Necessarily, all panelists addressed the need for AI governance, and shared top priorities on successful implementation of AI.

Components of good governance, the panel noted, should have these characteristics: Be consensus oriented, participatory, accountable, transparent, responsive, equitable and inclusive, effective and efficient, and follow the rule of law.

Review of New Developments and Future Impacts

Continuing the AI panel was Walter F. Wiggins, MD, a board certified neuroradiologist, Assistant Professor at Duke Health and Clinical Director of the Duke Center for AI in Radiology (DAIR) at Duke University School of Medicine. A strategic advisor to Qure.ai, Wiggins is focused on the use of advanced imaging and image analysis technologies in diagnostic imaging of the brain, head, neck, and spine, with a particular focus on the clinical implementation of artificial intelligence technologies for
medical imaging.

During his session segment, Wiggins discussed the topic of “Current Applications of AI in Radiology: What’s Out There? Where Are We Going?” He reported on the National Institutes of Health (NIH) Data Sharing Policy and updates, which apply to NIH grants received after Jan. 25, 2023. A data management and sharing plan will be required for those moving forward with applying AI in radiology, said Wiggins. He also identified the policy goals: Enable validation of research results; provide access to high-value datasets; and promote the use of data for future research. He offered that it is likely there will be continued growth in model/tool development and FDA clearance. Wiggins also suggested that mitigating bias and monitoring deployed tools are necessary components of the market strategy moving forward, and said analysis and determinations of who should bear that burden will be ongoing.

Wiggins addressed current applications and trends, noting the vendor landscape at RSNA 2022 showed a clear and consistent rise in both the number of companies in this space and the number of FDA-cleared AI tools.

He noted the recent release by the American College of Radiology (ACR) of an AI Central Dashboard. It can be broken down by subspecialty, by modality and allows users to track trends in clearance sequences. The dashboard also offers a catalog of all FDA cleared Software as a Medical Device (SaMD), of which currently there are 201, marking exponential growth from 2017 which is expected to continue.

Identifying other recent developments, Wiggins reported that chest radiographs continue to get a lot of attention, and noted a surge in text analysis tools, or natural language processing (NLP). He emphasized the heightened awareness and concern around bias — identification, mitigation and management. Citing a study in the journal Lancet published in early 2022 by collaborators from numerous institutions around the world, he suggested there will be more focus on training models in order to de-bias them in the future.

“If you want to see where things are going in the future, look to the FDA and how their guidance is coming out, as what we see there primarily is an increased emphasis on AI monitoring,” said Wiggins. “Then you also need to look at the NIH and what they’re funding, and what they’re saying you have to do once you get funding,” he added. Wiggins continued, saying: “The FDA is starting to move towards this concept of the total product lifecycle around the machine learning model, where potentially we can see the move towards continuous learning models, making it easier for vendors to update models and not have to go back through the FDA clearance process every time … this is still pretty preliminary, so we’re not sure exactly how this is going to play out. What I think we’re really going to see emerge from this, even before it gets things into the FDA, is an increased emphasis on clinical validation and post-deployment monitoring.”

He also noted the NIH data sharing policy, which went into effect in January 2023 saying, “What I think this is going to do to for AI research is to enable validation of research results, and provide access to high value data sets for reuse of data for future research, as those are the goals that NIH laid out.”

In closing, Wiggins forecasted continued growth in AI, as well as ongoing, important research for clinical implementation, and a focus on mitigating bias with the AI tools as necessary components of market strategy for vendors in the coming years. He returned to a shared concern moving forward, “The real question here is who is going to bear the burden? Is it vendors? Is it going to be us radiologists, or is it going to be a combined effort to try to monitor these tools?”