AI & Tech··1 sources verified
Vision-Language Model Agents Trail Radiologists Badly on Emergency CT and MRI Reads
Summarised by Relevant News AI · Read time: 3 min
In a retrospective study of 4,800 runs, six vision-language models read 100 emergency CT and MRI exams under four navigation conditions. Accuracy was lowest with free navigation (32.6%) and highest when all or key slices were supplied in one request (about 42%), while nine radiologists reached 77% on the 30 cases they read. The models typically committed to an early hypothesis, rarely recovered from a wrong one, and stated a median confidence of 90% even though only 38.9% of final answers were correct.
Why it matters: It gives practitioners concrete evidence that agentic VLMs overstate their diagnostic confidence and anchor on early guesses, which matters for any deployment that navigates multi-slice imaging.
All sources