The specialty’s AI journey has covered many miles in the decade since an infamous prediction of the demise of radiologists.
FROM THE CHAIR OF THE BOARD OF CHANCELLORS
Christoph Wald, MD, PhD, MBA, FACR
By Tessa Cook, MD, PhD, FACR
Chair of ACR’s Commission on Informatics
Guest Columnist
The modern version of the Hippocratic Oath recited by new medical students does not mention artificial intelligence (AI), imaging informatics, or even technology in general. And yet, these future physicians — some of whom will become our fellow radiologists — are entering a world of medicine that our predecessors never imagined.
Radiology has often found itself at the forefront of technological adaptation. With the advent of AI, we once again find ourselves leading the charge in adapting to this new technology. In June, the FDA updated its list of cleared software-as-a-medical-device (SaMD) AI tools, with nearly 80% of the cleared options representing radiology use cases.
Radiology’s AI journey has covered many miles in the decade since Geoffrey Hinton’s now infamous prediction of radiologists’ demise. We spent the first few years trying to understand how AI worked and what it could do. Then we began using narrow/specialist AI triage tools in our workflow — to prioritize cases with intracranial hemorrhage, pulmonary embolism and acute stroke. With time, AI expanded to include segmentation and quantification of organs and lesions, and opportunistic screening to quantify bone mineral density, atherosclerotic calcifications, hepatic steatosis and more. With the ubiquity of large language models (LLMs) that can generate report impressions, and vision-language models (VLMs) and foundation models that can produce entire draft reports for our review, we find ourselves on the precipice of one of the most dramatic innovations in our workflow since the introduction of PACS and voice recognition.
But AI is not like other technology we use. Clicking a button in the electronic health record has a behavior associated with it that is defined and predictable. On the other hand, AI is probabilistic, not deterministic, so its behavior can vary — sometimes subtly and unpredictably. That’s why it’s important to critically assess AI outputs and monitor them at scale.
We spent the first few years trying to understand how AI worked and what it could do. Then we began using narrow/specialist AI triage tools in our workflow — to prioritize cases with intracranial hemorrhage, pulmonary embolism and acute stroke.
Considering this, you won’t be surprised to learn that the ACR’s Commission on Informatics and the Data Science Institute® (DSI) have been busy. The first-ever practice parameter for imaging AI was developed last year and adopted by ACR Council in May, thanks to the tireless efforts of the writing and comment reconciliation committees. In addition, Council approved the exploration and development of AI accreditation.
At least 80% of FDA-cleared medical AI is for radiology, and nearly 90% of surveyed ACR members report using some form of AI in practice. But we know that real-world AI performance doesn’t always mirror what is observed during pre-market testing. And how many of us truly understand what our models are doing across our practices? For this reason, we need quality management (QM) tools to help us. ACR’s accreditation programs lead our specialty’s QM efforts, and a committee is actively working on developing accreditation for imaging AI. More than 37,000 modalities and practices have earned ACR accreditation since the program’s inception, and soon practices will be able to demonstrate their commitment to safe and effective patient care augmented by imaging AI.
One notable tool in the QM armamentarium is ACR’s Assess-AI service. Assess-AI provides real-world monitoring across dozens of use cases to better help us learn how models are performing and identify factors that might contribute to other-than-expected performance. A recent article in the JACR describes how Assess-AI combines data from the radiologist’s report, the DICOM header, and the AI output to produce a collection of metrics of model performance. These metrics guide participating sites as they deploy new models, monitor existing models and make decisions about whether to take models offline, sunset them, or update them based on observed performance.
Not sure where to start? Here’s a list of commonly expressed concerns by practices on the AI journey, along with the ACR DSI resources that can help:
ACR Council’s recent AI-related activities are not just paperwork and rubber stamps. They are concrete steps and real-world tools to support our members as they tackle the real and rapidly growing challenge of implementing AI in their practices — in a way that is safe, effective and moves the needle not only for our patients’ care but for our workforce experience.
These ongoing developments demonstrate how well ACR is positioned to support and oversee the deployment and monitoring of imaging AI. With DSI tools, resources, and programs for practices at all stages in their AI adoption, the first-ever imaging AI practice parameter, and the nascent AI accreditation program — ACR is ready to support members and practices in the safe, effective and impactful application of AI for patient care.
Sustainability in Radiology
The ACR Committee on Sustainability is leading the way in educating radiologists on how to prepare their practices to be more environmentally friendly.
Read more
Site-Neutral Imaging Payments
The Medicare Hospital CY 2027 OPPS proposed rule includes key policy changes that radiologists need to know about.
Read more
Continuing the Investment in Young Leaders
The 2026 cohort of the Philips Emerging Leaders RLI Scholarship builds on the success of its first year to prepare young leaders for the next generation of healthcare.
Read more
