Evolving Medical Imaging Agents via Experience-driven Self-skill Discovery

A research team led by Lin Fan and six other authors has introduced MACRO, a groundbreaking medical imaging AI agent that evolves its own capabilities through experience-driven self-skill discovery. Unlike conventional LLM-based agents that rely on static, predefined tool sets, MACRO autonomously identifies effective multi-step tool sequences from verified execution trajectories and synthesizes them into reusable composite tools. These new high-level primitives continuously expand the agent's behavioral repertoire, enabling it to adapt to real-world domain shifts, evolving diagnostic requirements, and diverse clinical tasks where traditional tool chains typically degrade.

The system employs a lightweight image-feature memory to ground tool selection in visual-clinical context, combined with a GRPO-like training loop that reinforces reliable invocation of discovered composite tools. This creates a closed-loop self-improvement mechanism requiring minimal human supervision. Extensive experiments across diverse medical imaging datasets demonstrate that autonomous composite tool discovery consistently improves multi-step orchestration accuracy and cross-domain generalization over state-of-the-art agentic methods. The approach represents a significant shift from brittle static tool composition toward adaptive, context-aware clinical AI assistance that can evolve alongside medical practice.

MACRO addresses a fundamental limitation in current medical AI systems: their inability to adapt tool usage strategies after deployment. By enabling agents to learn from their own successful execution patterns and formalize them as new capabilities, the system bridges the gap between rigid automation and flexible clinical reasoning. The research shows particular promise for complex diagnostic workflows where clinicians typically combine multiple specialized procedures, visual evidence, and patient context in iterative sequences that existing AI systems struggle to orchestrate effectively.