New medical imaging benchmark reveals that advanced AI models struggle with real-world diagnostic tasks requiring dynamic navigation of full 3D medical scans.

Researchers introduced MedOpenClaw, an auditable runtime enabling vision-language models to operate within standard medical imaging tools like 3D Slicer for more realistic diagnostic scenarios

MedFlowBench benchmark covers multi-sequence brain MRI and lung CT/PET studies, systematically evaluating AI agents across viewer-only, tool-use, and open-method tracks

Current evaluation methods oversimplify clinical reality by testing on pre-selected 2D images, missing the core challenge of navigating full 3D volumes across multiple modalities

Testing with state-of-the-art models including Gemini 3.1 Pro and GPT-5.4 revealed significant gaps in their ability to actively gather evidence from complete medical studies