This paper presents MINIM, a self-improving generative foundation model designed to work across multiple medical imaging modalities, including optical coherence tomography, fundus photography, chest X-ray, and chest CT. Unlike earlier generative models that are typically limited to a single organ or imaging type, MINIM is built as a unified system that can generate clinically meaningful medical images from text descriptions across diverse domains.
A defining feature of MINIM is its ability to improve after deployment. The model incorporates clinician feedback through reinforcement learning and uses transfer learning to adapt to new imaging domains, such as brain and breast MRI, without retraining from scratch. In practice, this means the AI can refine image quality and clinical relevance over time, moving closer to a generalist medical AI rather than a task-specific generator.
The study shows that MINIM’s synthetic images are not only visually convincing but functionally useful. When these generated images were added to training datasets, the performance of downstream AI systems—covering diagnostic classification, report generation, and self-supervised learning—improved by an average of 12% to 17% across ophthalmic, chest, brain, and breast imaging tasks. Objective image-quality metrics further showed that MINIM outperformed general-purpose generative models when evaluated in medical contexts.
Beyond data augmentation, the work demonstrates direct clinical relevance. Using chest CT images, the model was able to distinguish EGFR mutation status in lung cancer, separating treatment-sensitive from resistant cases, with image-based predictions associated with improved long-term survival. Similarly, in breast MRI, MINIM enabled accurate identification of HER2-positive tumors, supporting non-invasive selection of patients for targeted therapies.
Overall, this study reframes generative AI from a tool for image synthesis into a clinically adaptive system—one that can continuously improve, strengthen other AI models, and extract biologically meaningful information from routine imaging. The significance lies not in how the model is coded, but in how it learns across modalities, adapts with feedback, and supports clinical decision-making without additional invasive procedures.
J. Wang et al., “Self-improving generative foundation model for synthetic medical image generation and clinical applications,” Nat. Med., vol. 31, pp. 609–617, 2025, doi: 10.1038/s41591-024-03359-y
