MedGenAgent: a closed-loop agent for oncology CT foundation model pretraining

Oncology computed tomography (CT) foundation models require large, high-quality training datasets. However, the collection and annotation of such data are costly and time-consuming because tumors occur across multiple anatomical regions and have complex three-dimensional (3D) structures. Strict privacy constraints and institutional variation further limit the availability of large-scale oncology CT datasets. Here, we propose MedGenAgent, a closed-loop agent that supports oncology CT foundation model pretraining in data-constrained settings by iteratively constructing a lesion-rich, anatomically balanced 3D pretraining dataset. MedGenAgent incorporates a Specification resolver tool for structured task parsing and a central Orchestrator tool to coordinate the entire closed-loop agent workflow by leveraging both an Execution layer (including the MAISI synthesis model node for 3D oncology CT sample synthesis, the MedGemma vision-language model node for automated quality control, and the foundation model training node) and a Memory store tool for iterative, feedback-driven updates. Through the synergy of these tools, MedGenAgent synthesized 11,159 CT samples and retained 11,000 quality-controlled samples. By first leveraging this synthetic dataset for pretraining and then aligning with real-world data using only 4400 DeepLesion samples, representing 38.37% of the data used for baseline pretraining, the MedGenAgent-pretrained foundation model achieved performance comparable to the full-data baseline across three retrospective downstream tasks: lesion anatomical site classification, nodule malignancy classification, and cross-institutional NSCLC overall survival classification. Furthermore, automated quality control, independent physician assessment, and workflow analysis demonstrated the construction of a quality-controlled synthetic CT dataset while revealing regional variation in sample quality. Together, these results show that MedGenAgent can autonomously construct quality-controlled synthetic CT data that support the competitive performance of oncology CT foundation models when annotated real-world data are limited.