Current evaluations of vision-language models (VLMs) in medical imaging often oversimplify clinical reality. Relying on pre-selected 2D images, these setups demand significant manual curation and, critically, fail to address the core challenge of real-world diagnostics: a true clinical agent must actively navigate full 3D volumes across multiple sequences or modalities to gather evidence and support a final decision.
To tackle this, researchers have proposed MEDOPENCLAW, an auditable runtime designed to enable VLMs to operate dynamically within standard medical tools or viewers, such as 3D Slicer. This framework allows AI to move beyond passive image perception towards interactive, agentic behavior within clinical workflows.
Complementing MEDOPENCLAW, the team introduced MEDFLOWBENCH, a comprehensive full-study medical imaging benchmark. It covers multi-sequence brain MRI and lung CT/PET, systematically evaluating medical agentic capabilities across viewer-only, tool-use, and open-method tracks.
Initial results from MEDFLOWBENCH reveal a critical insight: while state-of-the-art LLMs/VLMs like Gemini 3.1 Pro and GPT-5.4 can successfully navigate viewers for basic study-level tasks, their performance paradoxically degrades when granted access to professional support tools. This degradation is attributed to a lack of precise spatial grounding, highlighting a significant limitation in current advanced AI models.
By bridging the gap between static-image perception and interactive clinical workflows, MEDOPENCLAW and MEDFLOWBENCH establish a reproducible foundation for developing auditable, full-study medical imaging agents. This advancement is crucial for fostering more practical, intelligent, and trustworthy AI solutions in healthcare diagnostics.