Physician liability when relying on AI recommendations; What is coming up at RSNA; Newly detecting LVOs with deep learning
Welcome to our November newsletter on neuroradiology and AI. We hope you find it helpful. Please feel free to share it with others.
If you would like to be informed when the new issue is out, please sign up here.
Conference Coverage: What to look for at RSNA
Conference updates: RSNA 2020 is all virtual, taking place 29 November – 5 December. The conference offers more than 350 sessions, courses and exhibits, plus 190 on-demand sessions. Look for sessions like: “Stroke Imaging — How Recent Trials Are Changing Radiologists’ Practices,” “Artificial Intelligence in Radiology: The Hype is Over — What is Next?” “Artificial Intelligence: Beyond Interpretive Considerations,” “Artificial Intelligence — Decision Support: The Coronavirus Experience in USA and China,” and “Artificial Intelligence and Implications for Health Equity: Will AI Improve Equity or Increase Disparities?” RSNA is offering other AI sessions on cardiology, oncology, gastroenterology, breast imaging, and informatics.
Neuroradiology and AI in the news
When does physician use of AI increase liability? This study, published in the Journal of Nuclear Medicine, notes that increasingly, AI and automated systems make medical treatment recommendations that can deviate from the standard care. The study questioned 2,000 US adults on the circumstances in which they would hold physicians liable for recommendations made by AI. Those surveyed were given four scenarios where the physician received an AI recommendation and either accepted or rejected it, and the recommendation was standard or nonstandard care – and all situations resulted in patient harm. Researchers found that when physicians accepted AI advice for standard care, those surveyed reduced the physicians’ liability compared to if the physician rejected the advice. If physicians rejected AI advice for nonstandard care, and provided standard care which was against the AI advice, survey takers would not shield them from liability. Researchers concluded that the legal system is not a barrier to adopting AI, but can actually encourage its use.
Detecting large vessel occlusion at multiphase CT angiography by using a deep convolutional neural network. As one of the most time-sensitive medical diagnoses, large vessel occlusion (LVO) requires emergency treatment to reduce morbidity and mortality. This retrospective study in Radiology focused on efforts to use deep learning to quickly detect LVO in multiphase CT angiography, reducing time to treatment. Researchers evaluated CT angiography examinations for 540 adults with suspected acute ischemic stroke. Of those, 270 were positive for LVO, confirmed by catheter angiography. The CT angiography preprocessing included vasculature segmentation and creating maximum intensity projection images to emphasize contrast agent-enhanced vasculature. Researchers performed experiments with various phases of the angiography. Their deep learning model could detect LVO, and the diagnostic performance was enhanced using delayed phases at multiphase CT angiography examinations.
Navigating the paradox of scarcity: The case for radiologist-driven care. Imaging utilization is rising while reimbursement is falling, leaving radiologists to determine how to manage their workload and attract new physicians to the field. Physician extenders are a popular option in primary care and in many specialties – including radiology. This opinion piece in JACR argues that instead of using physician extenders, radiologists should incorporate AI algorithms into daily practice. There are unintended consequences of allowing extenders to perform some radiology duties, which can potentially lead to missed incidental findings, for example. These are still the radiologists’ responsibility. The authors also fear a culture of “rubber-stamping” the extender’s findings without the radiologist properly reviewing them. Instead, the authors suggest that AI can be integrated into the workflow to improve efficiency.
AI-RADS: An artificial intelligence curriculum for residents. AI as a technology is not only gaining steam in many medical fields but of course in radiology as well. Dartmouth-Hitchcock Medical Center recognized the importance of AI in radiology and developed a curriculum to provide its residents with a background in the field, to better understand its foundations and language. The idea was to share the strengths and limitations of the technology and give an intellectual framework so the residents could evaluate the literature moving forward. The curriculum consisted of lectures and journal clubs. The course was well-received by residents, receiving high marks. The faculty shared their insights in Academic Radiology.
Bilateral distance partition of periventricular and deep white matter hyperintensities: Performance of the method in the aging brain. Researchers developed an AI calculator to analyze brain tissue damage seen on MRI scans, to detect early signs of cognitive decline, yielding a 70% accuracy rate. A study published in Academic Radiology shared the development of an algorithm that calculates sizing of white-matter hyperintensities as an additional standardized test to assess these brain lesions, in advance of stroke damage or severe dementia. The new calculator can determine lesion volumes based on distance from each side of the brain. The algorithm’s measurements compared against the MRI findings were 70% accurate. The researchers will complete additional testing with almost 1,500 more MRI scans.
AI continues to make in-roads in many parts of medicine. In the brain, algorithms to identify abnormalities can lead to faster treatment and better outcomes. Legal implications for relying on AI are less concerning than some people might think, and patients have some faith that the algorithms and their recommendations are valid. Teaching the basics of AI to radiology residents can move the field forward and increase acceptance. When provided with an educational foundation, the upcoming medical community can experience greater growth and have more informed discussions of AI’s development and use. While AI will not replace radiologists, it can make the workflow more efficient, allowing radiologists to focus on more complicated cases.