RSNA president offers 6 ingredients for a ‘better AI future’

CHICAGO—Curtis P. Langlotz, MD, PhD, a noted physician and president of the Radiological Society of North America, is offering up six ingredients that could make for a “better AI future” in imaging. 

He spoke Sunday night during the President’s Address and opening session of RSNA’s 110th Scientific Assembly and Annual Meeting.  Langlotz, a professor of radiology at Stanford, chose “building intelligent connections” as the theme of his talk. 

“As technology has advanced in recent years, imaging has become central to patient care, but many radiologists still work in relative isolation with few breaks,” Langlotz told a packed Arie Crown Theater at McCormick Place on Dec. 1. “If we want to thrive as a specialty, we need to build a new foundation for our practice, one that's based on connections—connections with other healthcare practitioners to improve patient care, connections between new technologies and systems for better workflow and connections with each other in the broader world to exchange ideas and to innovate.”

As an example, Langlotz gave the story of a 55-year-old former smoker named Lane who had hemolytic-uremic syndrome as a child. In 2015, the woman fell from her bike and was brought to the ED with sharp, central and left-lateral chest pain. Her EKG was normal, and she underwent a contrast-enhanced CT of the chest, which was interpreted as showing no evidence of aortic or other traumatic injury. 

Because of her history of chronic renal failure, Lane experienced contrast-induced acute kidney injury and needed to be observed in the hospital overnight. An older detection system, typical of nine years ago, flagged four possible nodules. A radiologist reviewed each alert and decided only one was worthy of follow-up, based on Fleischner Society guidelines, a 7-centimeter solid nodule. However, in the ensuing days, Lane continued to experience severe left-lateral chest plan. Follow-up X-rays five days later detected minimally displaced fractures of the ribs. And six months later, she underwent follow-up CT at a facility near her home to address the lung nodule. 

But the prior images were unavailable, and a biopsy was recommended. 

“Consider what Lane must do to retrieve the prior images,” Langlotz told attendees. “She travels to the hospital that treated her after the bike accident, pays a $10 fee to have a CD burned with her images and delivers the CD to the imaging center,” showing the nodule was unchanged and leading to the biopsy being canceled. 

The RSNA president asked attendees to imagine how the 2015 scenario would have gone in a more connected healthcare landscape. After the bike accident brought Lane to the ED, an AI-generated patient summary of her health information would alert the care team that she has a history of renal insufficiency. IV isotonic volume expansion with normal saline is employed, dramatically reducing her risk of renal injury and preventing a hospital admission. Providers deploy a suite of AI models capable of detecting all relevant findings on her trauma CT. The software identifies the two rib fractures and a single solitary nodule that needs follow-up. 

Because newer machine learn models are more accurate, there are no false-positive nodule alerts. These findings are then automatically inserted into a structured report for the radiologist to review. A decision-support system alerts the ED physician to the findings and Lane is given instructions for rib pain management and nodule follow-up. In the comfort of her home, Lane interacts with a chatbot trained on high-quality medical data, which explains the radiology report in simple language. And when her nodule is imaged at the facility near her home, the PACS automatically obtains her prior images from a national exchange, enabling a direct comparison to show that the nodule remains unchanged. 

“This new approach has significant benefits for Lane and patients like her,” Langoltz said. “These same enhancements also benefit radiologists and the entire healthcare system through improved deficiency, better patient care, and reduced malpractice risk. So, the question is how do we get to this better world, a world enhanced by AI?”

Langoltz offered six ingredients to reach this ideal state: 

1. Ditch the disk: “First, good care starts with good data. The CD is almost as old as the fax machine and just as useless. We shouldn't be putting our sick patients through the hassle and cost of carrying their data between facilities. So, let's ditch the disc. We need to urge our image-exchange vendors to enable universal electronic image exchange. Many healthcare organizations already exchange information, medical record data over the internet, so why not exchange image information the same way?”

2. Make data donation easy: “Second, we should encourage sharing of data to train AI. Surveys show that most patients are happy to contribute their data to research once they understand the intention is to improve life for patients like them. A great start would be to enable patients to easily opt in to donate their de-identified data for research just like organ donation. We must consider patient and provider needs as we build these systems. As privacy advocates say, ‘nothing about me without me.’ That means having a diverse interdisciplinary team in the room from Day 1, participating in key design decisions and guiding the collection of training data to assure it will produce a fair AI model that provides accurate answers for everyone. And in this era of cyberattacks, we all must preserve the privacy of patient data by taking every security precaution and by de-identifying data we use for research.”

3. Improve human-machine collaboration: “We need a much better understanding of human-machine connections to avoid automation bias and other ways in which machines can degrade human performance. For example, AI models that explain their reasoning and express their confidence or the lack of it should be the norm.”

4. Modernize our regulatory framework: “Fourth, our regulatory framework needs to catch up with modern reality. Our current regulatory framework puts too much emphasis on the pre-market clearance process and not enough emphasis on what happens after models are deployed. Our regulators are doing a good job under difficult constraints, but they're laboring under a 50-year-old regulatory framework. It's like trying to run a 2024 road race on the chassis of a 1976 Oldsmobile Cutlass and today's AI customers are often flying blind.”

5. Share objective information about AI systems: “Model developers should publish model cards like a nutrition label for AI models to help radiologists decide whether a model will work in their practice…In reports of clinical trials, they provide information about the model's training. For example, many of today's AI models are trained on adult data but may be deployed in pediatric settings where they're unsuitable. Model cards will help us avoid these pitfalls, and we need a Consumer Reports for AI to provide unbiased comparisons of models creating competition between vendors on accuracy and quality.”

6: Monitor system performance over time: “Finally, AI should be rigorously evaluated to prove that these new tools are safe and effective using prospective clinical trials for algorithms deployed in high-stake settings. After system implementation, AI and quality improvement experts at each site must work together to track adverse events and assure the system remains accurate as the data shifts, such as when a new device is brought online, a clinic opens with a different patient mix or a new strain of viral pneumonia spreads among the population.”

“In summary, this prescription for an AI future requires universal internet-based image exchange, easy ways for patients to donate their data to research, more research on human computer interaction, modernized AI regulations and more objective information about AI algorithms as well as system performance monitoring over time,” Langlotz added. “We too often think of high tech and high touch as opposite ends of a spectrum. High touch often signifies human-centered technologies to build human connections and improve our lives and our careers, and we typically think of high-tech systems as complicating our work and drawing us away from real world human connections. But they're not opposites. The latest high-tech advances are likely to give us more opportunities for high-touch care. The latest wave of high tech will allow us to delegate the least attractive parts of our jobs to the AI models while retaining the rewarding tasks, focusing on patient care and personal connections. These advances can upskill us all, reduce burnout and bring better healthcare to underserved areas, and it can do so while we develop richer human connections like the ones we form in the reading room, in the exam room, and at meetings like this one.”

Other news

Meanwhile, in other RSNA news, the society named: 

• Umar Mahmood, MD, PhD, as president of the Radiological Society of North America Board of Directors on Sunday. A radiologist at Massachusetts General Hospital in Boston, Dr. Mahmood serves as chief of Nuclear Medicine and Molecular Imaging, where he oversees a service that spans multiple hospitals and facilities in the region. 
• Tina Young Poussaint, MD, as member of the RSNA Board and liaison for publications. She is the radiologist-in-chief and the Lionel W. Young Chair of the Department of Radiology at Boston Children’s Hospital and a professor of radiology at the Harvard. 
• Jeffrey S. Klein, MD, as chair of the RSNA Board. A renowned expert in lung cancer staging and detection, he is the A. Bradley Soule and John P. Tampas Green and Gold Professor of Radiology at the University of Vermont College of Medicine, in Burlington, Vermont.
• Anne Covey, MD, as a member of the RSNA Board and liaison for public information and professionalism. She is an interventional radiologist and professor of radiology in New York City at Weill Cornell Medical College and an attending radiologist at Memorial Sloan Kettering Cancer Center, where she has practiced since 2000.