AI Agents in Radiology: A Step Towards Autonomy

Modern radiology relies on complex workflows, where clinicians combine visual image analysis with patient clinical data, quantifying results through specialized procedures. LLM-based agents promise to orchestrate these heterogeneous tools, but current systems tend to treat tools and usage strategies statically.

MACRO: Continuous Learning for Medical Imaging

The proposed solution, called MACRO, introduces a self-evolving medical agent that shifts from static tool composition to experience-driven tool discovery. MACRO autonomously identifies effective multi-step tool sequences, synthesizes them into reusable composite tools, and registers them as new high-level primitives, continuously expanding its behavioral repertoire.

Implementation Details

A lightweight memory system based on features extracted from images guides tool selection in the visual-clinical context. A training loop reinforces the reliable use of discovered composites, enabling closed-loop self-improvement with minimal supervision. Tests on various medical imaging datasets demonstrate that autonomous composite tool discovery improves multi-step orchestration accuracy and cross-domain generalization compared to existing methods.