AI
Diagnostic Accuracy and Navigation Behavior of Vision-Language Model Agents Reading Emergency CT and MRI Examinations
Publication: cs.AI updates on arXiv.org
Author: Jonathan Lichtle, Aymen Meddeb, Antonin Birot, Ibrahima Niang, Lisa Gastaud, Riyad Hanafi, Sofiane Dujardin, Corentin Selomme, Guillaume Cachard, Aghiles Hamroun, Gr\'egory Kuchcinski, Su Hwan Kim, Ba
Source published: 19/11/2026, 05:00:00 UTC
Briefing created: 19/11/2026, 05:10:56 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 19/11/2026, 05:03:24 UTC.
Briefings
A retrospective study evaluated six vision-language model agents' diagnostic accuracy and navigation behavior on 100 emergency CT and MRI cases from March 2025 to August 2026. VLMs performed significantly worse than human radiologists, achieving only 30% accuracy on free navigation versus 77% for radiologists. Models frequently adopted incorrect early hypotheses and rarely corrected them, resulting in low confidence despite high stated confidence levels.
Key points
- Vision-language model agents achieved 30% accuracy on free navigation tasks, significantly lower than human radiologists' 77% accuracy on the same emergency cases.
- Models adopted incorrect early hypotheses in 87.6% of runs and failed to correct them, resulting in a 90.4% success rate only when the initial hypothesis was correct.
Context
This study investigates how vision-language models perform in emergency radiology, comparing their navigation strategies against human radiologists. The findings are critical for understanding the reliability of AI-assisted diagnostics in time-sensitive clinical settings.
Limitations
Data covers only March 2025 to August 2026, with 100 cases and 30 radiologist reads. Generalized estimating equations were used for statistical analysis, and the study relies on a single public archive with no external validation.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 19/11/2026, 05:10:56 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeigylolgalk4qcwtwjrjwsjkd5r27ayhwbnlztr7fvpudvu4nodlsu