Postdeployment Monitoring and Surveillance Methods, Guidelines, and Possibilities for AI in Radiology
Radiographics, 2026
As radiology AI tools move from pilot projects into daily clinical use, the real test isn't whether they work at launch, it's whether they keep working reliably once deployed. This new review in tackles exactly this problem: how should hospitals and imaging practices actually monitor AI once it's live?
The authors lay out a spectrum of human oversight models, ranging from having a radiologist check every single AI output, to fully autonomous operation with no human check at all. They argue for a middle-ground approach called "human-on-the-loop," where the system runs independently but automatically flags concerning changes for human review, rather than requiring constant manual checking. The clever part is how it detects problems without needing instant ground-truth answers: by tracking whether the AI's inputs and outputs stay stable over time, whether its predictions are drifting from baseline, and how confident it is in each output. When something looks off, a tiered alert system kicks in, escalating from a quiet note in a log all the way up to pausing the AI for a specific scanner or use case while the issue gets investigated.
The review also maps out how this fits global regulatory expectations, from the FDA to the EU AI Act to authorities in Singapore, Brazil, and beyond, all of which increasingly expect this kind of ongoing oversight rather than a one-time validation.
It's a practical, structured framework for anyone building or evaluating governance for clinical AI tools, useful reading whether you're a radiologist, hospital IT lead, or vendor.