Eleftherios Tzanis, Michail E. Klontzas
Objective To develop an autonomous, open-source, multi-agent system capable of performing end-to-end artificial intelligence (AI) workflows in medical imaging, including exploratory data analysis, feature importance analysis, radiomic feature extraction, and the development and deployment of segmentation, classification, and regression models. Materials and Methods The system comprises a master agent that coordinates multiple task-specific agents, each responsible for executing a distinct AI function. Based on natural language prompts, the system performs reasoning, selects the appropriate tools, and carries out complex, multi-step workflows autonomously. The system's performance was evaluated using 16 open-source datasets spanning structured clinical data and multimodal medical imaging across a range of anatomical regions and pathologies. Evaluation included a diverse set of task and workflow prompts. Output correctness was confirmed by monitoring system logs for agent and tool behavior and by re-executing all tools manually to ensure identical results. Results The system successfully executed all tasks when powered by high-end large language models (LLMs), demonstrating robust performance across exploratory and feature importance analysis, radiomic extraction, segmentation, and predictive modeling. At no point did the user instruct mAIstro which task-specific agents to invoke or how to solve the task. mAIstro autonomously understood the prompt, devised a strategy, selected the appropriate agents, and coordinated their execution to complete complex workflows without human intervention. All outputs matched those produced through manual tool execution. Conclusion This study introduces an LLM-agnostic, multi-agent framework for reproducible and autonomous development of medical AI pipelines. Code and implementation instructions are available at: https://github.com/eltzanis/mAIstro.