AI Costs vs. Validation

A radiology group can now buy an artificial intelligence read for about the price of a vending-machine soda, while the journal articles landing on radiologists’ desks are saying the real work starts after the software goes live. Zebra Medical Vision was already advertising image analysis for $1 per scan in 2017, and that low-friction pricing model never really went away. (auntminnie.com) Radiology is a natural market for artificial intelligence because a scan is already a digital file, and software is good at sorting through piles of digital files. The United States Food and Drug Administration said last week that its public list exists to track artificial-intelligence-enabled medical devices already authorized for marketing in the United States. (fda.gov) That flood of products is heavily concentrated in imaging. A 2025 Radiology best-practice statement said that, as of May 2024, 671 of 882 United States Food and Drug Administration-cleared artificial-intelligence-enabled devices, or 76.1%, were for radiology practice. (pubs.rsna.org) Most of these tools are not replacing the radiologist reading the scan from scratch. They are doing narrower jobs like flagging a possible lung nodule, moving a suspected emergency case higher in the worklist, or measuring something tedious that a human could miss at the end of a long shift. (pubs.rsna.org) That narrow role is why the pricing can look so cheap. If a vendor charges per scan instead of selling a giant enterprise contract, an imaging center can treat the software like an extra pair of eyes added to an existing workflow instead of a new department with a new payroll line. (radiologybusiness.com) The catch is that a low sticker price does not mean a low evidence burden. An American Journal of Roentgenology paper published online on March 4, 2026 said real-world performance of radiology artificial intelligence applications “frequently diverges” from previously reported results, based on a prospective evaluation of 13 models across 88,645 examinations. (ajronline.org) That gap shows up because hospitals are messier than benchmark datasets. The same March 2026 paper tested models across clinical sites from July 2022 to November 2024, and the whole point was to predict value before deployment because published accuracy numbers alone were not enough. (ajronline.org) Radiology journals are now saying hospitals should not treat clearance by the regulator like the end of the story. A 2025 American Journal of Roentgenology clinical perspective said successful deployment needs rigorous validation, real-world testing, continuous feedback, and robust monitoring processes after the tool is installed. (ajronline.org) The specialty’s own best-practice statements are moving the same way. The 2025 Radiology guidance for thoracic imaging said practical implementation requires objective on-site performance evaluation and postdeployment monitoring, which means checking how the tool behaves on your scanners, your patients, and your reporting workflow rather than trusting a vendor slide deck. (pubs.rsna.org) Regulators are moving in that direction too. The United States Food and Drug Administration opened a public request in 2025 on how to measure real-world performance of artificial-intelligence-enabled medical devices and how to detect performance changes over time after deployment. (fda.gov) So the market is settling into a two-track reality. Buying the software can look cheap and easy on a per-patient basis, but using it responsibly is starting to look more like running a quality-assurance program that never really ends. (fda.gov)