MedOpenClaw: Auditable Medical Imaging Agents Reasoning over Uncurated Full Studies

MedOpenClaw provides an auditable runtime that lets vision-language models interact with standard medical viewers to navigate full 3D studies.

MEDFLOWBENCH is a new full-study benchmark that covers multi-sequence brain MRI and lung CT/PET and tests agents across three tracks.

Initial evaluations show top LLMs/VLMs can solve basic study-level tasks via viewer navigation but perform worse when given professional tools due to poor spatial grounding.

The work aims to close the gap between static-image evaluation and interactive clinical workflows by offering a reproducible platform for auditing medical imaging agents.