Towards Autonomous and Auditable Medical Imaging Model Development

AMID, from researchers at The Chinese University of Hong Kong, the Chinese Academy of Sciences, Lehigh University, and Microsoft Research, is an autonomous multi-agent system for building auditable medical-imaging models. Its first contribution, Data-Conditioned Method Planning, profiles modality, geometry, labels, patient organization, metrics, and submission requirements, then converts them into executable method lanes grounded in resources such as nnU-Net, MONAI, MedSAM, and pathology encoders UNI and Phikon-v2. Its second contribution, Verification-Guided Two-Stage Optimization, uses broad behavior-gated exploration followed by selective exploitation, while independent reviewers verify data splits, metric calculations, prediction schemas, checkpoints, and traceability before evidence can influence promotion. On the ReX-MLE benchmark’s 20 medical-imaging challenges, using a 24-hour budget, one NVIDIA RTX A6000 GPU, and Codex with GPT-5.5, AMID produced accepted results on every task and improved over the strongest listed autonomous baseline on 19 tasks. It achieved 0.91 Dice on SEG.A, a 0.89 absolute gain over the best baseline, and 0.49 AP on DENTEX, matching the human reference; task-specific gains came from combining YOLOv8 with anatomical calibration and using foundation-model features for pathology. Focused experiments also used Claude Code, which reached 0.64 Dice on PUMA-T1-Seg. The report emphasizes that AMID’s advantage comes from organizing and auditing the search process, not merely from the underlying coding agent, while noting unresolved costs, backend variability, and the absence of clinical validation.