AI: capability and risk

Medical artificial intelligence is moving from pilot projects into daily radiology work, but doctors and researchers say its safety still depends on the data it learns from and the controls around it. (wgbh.org) A radiology AI system usually learns by finding patterns in huge sets of scans and reports, like a resident who studies old cases before reading new ones. Marzyeh Ghassemi, an Massachusetts Institute of Technology professor, told GBH on April 22 that biased or incomplete training data can push those systems toward biased or unsafe results. (wgbh.org) The National Institutes of Health reported in July 2024 that one image-reading AI model often chose the right diagnosis but still made mistakes when describing the image and explaining its reasoning. In that study, physicians using outside resources outperformed the model on the hardest cases. (nih.gov) Radiology groups are now building AI for workflow as much as diagnosis. The Radiological Society of North America says it funds peer-reviewed research, develops “ground-truth” datasets, and publishes implementation guidance aimed at practical and ethical use. (rsna.org) One current target is follow-up care after a scan spots something that needs another test. A Journal of the American College of Radiology review published in late 2025 broke that job into four steps: finding the recommendation in the report, communicating it, tracking completion, and measuring outcomes. (jacr.org) RSNA highlighted one such project on April 21, 2026: an alert system for inferior vena cava, or IVC, filters that scans finalized computed tomography reports and flags patients who need follow-up. The society said the tool detected IVC filters in 99.7% of reports, with 85.7% sensitivity, 99.9% specificity, and a 94.7% positive predictive value. (rsna.org) Vendors are also stitching separate tools into larger operating systems. Deepc and DeepTek said on April 8, 2026 that they had integrated deepcOS and DeepTek’s Augmento platform so hospitals could deploy, monitor, and govern multiple AI products through one environment. (deepc.ai) That push toward scale is colliding with a newer threat: fake medical images that look real enough to fool experts. In a Radiology study released by RSNA on March 24, 2026, 17 radiologists from 12 centers in six countries reviewed 264 X-rays, half real and half AI-generated. (rsna.org) When readers were not told the study involved synthetic images, only 41% spontaneously noticed anything unusual. After they were warned, their mean accuracy at separating real from fake X-rays was 75%, while four multimodal large language models scored between 57% and 85%. (rsna.org) RSNA said the study points to risks that go beyond misreads, including fraudulent legal claims and the possibility that hackers could insert fake images into hospital systems. Its 2025 statement on radiology AI also calls for validation, monitoring, transparency, and human oversight as these tools move into practice. (rsna.org, rsna.org) The pattern across these projects is that radiology departments are buying less of a single algorithm and more of a chain of systems: data curation, model testing, workflow tracking, and image verification. As more scans, reports, and alerts pass through AI, the question is no longer only whether a model works in a lab, but whether a hospital can prove where an image came from and who checked the output before it reached a patient. (rsna.org, deepc.ai, wgbh.org)