Our Take
The FDA is clearing generative AI faster than the field can validate it, which means early adopters will become the validation studies.
Why it matters
Radiologists spend significant time on report drafting. If these tools work, they free capacity for interpretation work that AI cannot yet do reliably. The regulatory precedent matters more than the products: the FDA is bending its framework to fit vision-language models, not waiting for it to catch up.
Do this week
Radiology leaders: document your current report turnaround time and error rates this week so you can measure whether breakthrough-designated tools actually compress clinical workflow or just shift the review burden.
Two AI tools win FDA breakthrough status
The FDA granted breakthrough designation to Cognita (acquired by Radiology Partners in late 2025) in March, and to Aidoc's First Read tool on Thursday. Both use generative AI to interpret chest X-rays and draft the radiology reports that radiologists typically write by hand.
The technical shift is material. Machine learning systems have long flagged suspicious regions in images for human review. Generative AI does something different: it processes the entire image and outputs prose findings for a radiologist to review and sign. This is not annotation; it is report generation.
Aidoc's designation covers detection and description of four life-threatening findings. The breadth of Cognita's scope was not disclosed in available reporting.
Regulatory speed outpaces validation method
Breakthrough designation signals the FDA expects the device to offer substantial advantage over existing practice. For radiology, that logic holds: report writing is a bottleneck, not a diagnostic bottleneck. If generative AI can draft accurate findings, radiologist time shifts from typing to interpretation.
But the FDA's usual validation toolkit does not fit. Traditional machine learning on imaging is evaluated on held-out test sets: sensitivity, specificity, area under the curve. Generative AI on images produces natural language. Evaluating report quality requires radiologist review of prose, not pixel-level metrics. The FDA acknowledged this gap: the article notes validation "is a challenge" and the process is "challenging traditional validation and regulatory frameworks."
This is not new regulatory risk. It is unresolved regulatory risk. The breakthrough designations mean two things: first, the FDA believes the technology has merit; second, the FDA is willing to approve it before the field has settled on how to prove it works. Early adopters will be the validation cohort.
Measure before you deploy
Report generation time and accuracy are not the same thing. A tool that drafts reports faster but introduces errors that radiologists must catch and correct does not improve throughput. It redistributes work.
Before piloting either tool, establish baseline metrics on your current radiology practice: average report turnaround time, rate of amendment or correction within 24 hours, and radiologist time spent typing versus interpreting. Deploy in a controlled cohort. Measure the same metrics under the AI tool. Measure radiologist satisfaction separately.
The FDA's breakthrough designation means the tool passed an efficiency or safety bar relative to status quo. It does not mean it will improve your specific department's workflow. That is a local question with a local answer.