Image source: Shutterstock/LeoWolfert
The keynote address of the 2021 Society of Imaging Informatics in Medicine (SIIM) annual meeting, the Samuel Dwyer Memorial Lecture, held virtually in late May, focused on the exponential development of medical AI research and “approved” radiology AI tools, as well as the challenges to achieve expected quality, safety, and performance attributes.
Adam E. Flanders, M.D., Professor of Radiology and enterprise vice-chair for imaging informatics at Thomas Jefferson University Hospital in Philadelphia, PA, delivered the lecture, entitled “Serendipity and Unintended Consequences”. It reflected the visionary perspective of the late PACS pioneer Professor Samuel J Dwyer, III, Ph.D., whose seminal research on the retrieval rates of radiology exams formed the basis for PACS storage systems. Dr. Flanders likened the surge in AI tool development and image data set expansion to the space race of the 1960’s, whose serendipitous and unintended consequence was the launch of the digital age, and to the global campaign to end the use of the harmful pesticide DDT. The latter produced a negative unintended consequence: injury to wildlife was reduced, including mosquitoes, which then caused a large increase in cases of malaria.
The promise of AI technology applied to radiology is huge. It has the potential to improve diagnostic quality, to efficiently extend access, and to be a panacea to major cost reductions in medicine. Dr. Flanders referenced experts who predict that healthcare AI overall will be a US $36 billion industry by 2025, generating a US $150 billion annual savings in AI-enabled healthcare by 2026. HealthITAnalytics reported in 2019 that the top three areas of healthcare AI investment were robot-assisted surgery ($40 billion), virtual nursing assistants ($20 billion), and administrative workflow ($18 billion). By comparison, AI for preliminary diagnosis and automated image diagnosis ranked at the bottom of its list, at $5 billion and $3 billion respectively.
AI needs guardrails. It doesn’t know when it doesn’t have enough information when making a decisionAdam E. Flanders
Nonetheless, there has been a tremendous surge in interest in AI in radiology, stimulated in part by the 2016 prediction by Geoffrey Hinton, Ph.D., one of the godfathers of deep learning and neural networks, that AI would diagnostically outperform radiologists by 2021 His statement was a wake-up call to radiologists that AI technology would have a huge impact on their medical specialty. Since then, there has been a surge in interest in the potential of medical imaging AI. “But is diagnostic AI on the right track? Is it happening too fast? Are image databases used to train AI software diverse enough, inclusive enough? Is there enough information?” queried Dr. Flanders. “AI needs guardrails. It doesn’t know when it doesn’t have enough information when making a decision.” Dr. Flanders cited an example of a brain CT angiogram whereby the AI interpretation was a technical error of insufficient contrast material, when, in fact, the patient was brain dead. While a radiologist can infer the correct diagnosis in this instance, it underscores the fact that AI is fundamentally limited to entities that it has been sufficiently trained upon.
As of 2021, 222 AI medical devices had been cleared by the U.S. Food and Drug Administration (FDA). Of these, 129 were for radiology applications, with 77% manufactured by small companies. But there is no agreed-upon definition of an AI device or specific regulatory pathway, either within the U.S. or in Europe. Is there enough rigorous evaluation? Dr. Flanders’ doesn’t think so, nor do many other radiologists starting to publicly speak out. Researchers from Radboud University Medical Center in Nijmegen, the Netherlands, evaluated 100 CE-marked AI software products for clinical radiology from 54 vendors listed in www.aiforradiology.com. They reported in European Radiology that 64 products had no peer-reviewed evidence of efficacy, and that only 18 of the 100 had evidence of efficacy rated as 3/6 or higher, validating impact on diagnostic decision-making, patient outcome or cost.
Image source: van Leeuwen et al., European Radiology 2021 (CC BY 4.0)
Between January and October 2020, 320 papers were published describing new machine learning-based models to detect Covid-19 in chest CT images and chest radiographs. Writing in Nature Machine Intelligence, researchers from the University of Cambridge reported that none of these models were of potential clinical use due to methodological flaws and/or underlying biases. Most failed to describe a reproducible methodology, or failed to follow best practice for ML model development, and/or failed to show sufficient external validation to justify the wider applicability of the method.
Covid-19 is the initial research use case, but there is every reason to believe that this infrastructure can be expanded to support medical image AI research for many other conditionsAdam E. Flanders
Dr. Flanders said that a recent editorial in Radiology entitled “Artificial Intelligence of Covid-19 Imaging: A Hammer in Search of a Nail” was even blunter. “To a large extent, the large quantity and rapid publication of articles on AI for Covid-19 are emblematic of current trends in other areas of radiology AI. It is now so much easier to design and conduct a radiology AI experiment. The only prerequisite seems to be possession of a large data set,” commented Ronald M. Summers, M.D., Ph.D., of the National Institutes of Health Clinical Center’s Imaging Biomarkers and Computer-Aided Diagnosis Laboratory of Bethesda, M.D. “International standards to enable meaningful comparisons of the performance are needed for AI software, emphasized Dr. Flanders. “Prospective outcomes studies are necessary to determine whether use of AI leads to changes in patient care, shortened hospitalizations, and reduced morbidity and mortality.”
Access to data is critical, but it has historically been monolithic, isolated, and difficult to access. Governments are now starting to support image data collaboration. Covid-19 has provided a huge opportunity, because significant R&D funds became available to facilitate image data sharing and collaboration. One example is the Medical Imaging and Data Resource Center (MIDRC), a consortium formed by the American College of Radiology (ACR), the Radiological society of North America (RSNA), and the American Association of Physics in Medicine (AAPM) for image collection, curation and distribution of Covid-19 related imaging data for AI research. MIDRC supports 12 internal Covid related research projects alone. Its repository is expected to be the largest open database of anonymized Covid-19 medical images and associated clinical data in the world. “There is not an organ system in the body that is not affected by Covid-19. I predict that this image repository, which welcomes image data contributions from throughout the world, will be of tremendous benefit to AI research and innovations for radiology,” Dr. Flanders commented to Healthcare in Europe. “Covid-19 is the initial research use case, but there is every reason to believe that this infrastructure can be expanded to support medical image AI research for many other conditions.”
With respect to continuing AI radiology research, Dr. Flanders’ advice is to contribute data to public data repositories to mitigate bias and strength diversity in AI research and development. He recommends that new international evaluation checklists for testing be promoted. Radiologists should retain an appropriate level of skepticism when assessing products for clinical use, and when purchasing radiology AI tools, mandate vendors to provide a complete account of performance. “Most importantly, remain engaged in this very exciting, accelerating, and ever-changing process. Just as radiologists’ input helped to perfect PACS to be the utilitarian technology it is today, most likely, they will do the same for radiology AI tools as they become ubiquitous,” he concluded.
Adam E. Flanders, M.D. is Professor of Radiology and Rehabilitation Medicine at Thomas Jefferson University Hospital in Philadelphia. He is co-director of neuroradiology/ENT radiology and vice chair of imaging informatics. In this capacity, he chairs the Imaging Informatics Council for the Enterprise Radiology and Imaging Service Line for Jefferson. Dr. Flanders has been actively involved in informatics-related activities for SIIM and the RSNA for much of his career. He has been leading a multi-center collaborative project that is examining imaging features, pathology, and genomics of human gliomas from the Cancer Genome Project using images contributed to the Cancer Imaging Archive (TCIA) as part o the NIH Cancer Imaging Program.