Synthetic medical data is rapidly transforming the landscape of healthcare by enhancing diagnostic accuracy, treatment planning, and research capabilities. The integration of synthetic data effectively addresses critical challenges in the medical field, particularly in areas where data is scarce or difficult to obtain. In our paper, we present a foundational model named MINIM, designed specifically for generating synthetic medical images. This model generates text-guided images across various medical imaging modalities, paving the way for improved clinical applications and research advancements.

Our team has focused on creating MINIM by integrating a diverse array of medical imaging datasets for training purposes. These datasets include Optical Coherence Tomography (OCT), chest X-rays, chest CT scans, and fundus images. Clinician assessments and objective image comparison metrics indicate that MINIM generates synthetic medical images that closely resemble real medical images, effectively illustrating the imaging characteristics of various diseases. This fidelity is crucial for ensuring that synthetic data can be reliably used in clinical settings.

By combining synthetic medical images with real-world images, we found that incorporating a specific proportion of synthetic data significantly enhances the performance of various downstream medical tasks. For instance, in tasks related to diagnosis based on medical imaging and image report generation, the inclusion of synthetic images not only improves performance metrics but also aids in reducing biases that may arise from limited datasets. This approach allows for a more robust training framework that can adapt to diverse clinical scenarios.

To facilitate self-iteration and continuous self-improvement within our model, we have implemented a Reinforcement Learning from Human Feedback (RLHF) strategy. This innovative approach leverages feedback from physicians who evaluate the quality of the medical images generated by MINIM. Clinicians score these images on a scale from 1 to 3 based on quality and alignment with the input texts, and these scores serve as a reward for further training our model. Experimental results demonstrate that this RLHF-based method is effective for training generative AI models, representing a significant technological advancement in artificial intelligence applications within medicine.

Our team has further applied synthetic data in clinical settings through a retrospective simulation analysis. This analysis highlighted MINIM's potential in accurately identifying targeted therapy-sensitive EGFR mutations in breast cancer MRI images. By improving diagnostic accuracy for these mutations, we aim to enhance treatment planning and increase five-year survival rates for patients. The ability to accurately detect these mutations is crucial for tailoring personalized treatment strategies, thereby improving patient outcomes.

In summary, our work with the MINIM model underscores the transformative potential of synthetic medical data in healthcare. By integrating diverse imaging datasets and employing a Reinforcement Learning from Human Feedback strategy, we have developed a model capable of generating high-quality synthetic medical images while enhancing diagnostic accuracy in clinical applications. The promising results, particularly in identifying targeted therapy-sensitive EGFR mutations in lung cancer and HER2 mutations in breast cancer, highlight the potential for improved treatment planning and better patient outcomes. This innovative approach paves the way for advancements in artificial intelligence within the medical field, ultimately contributing to enhanced patient care and operational efficiency in healthcare systems.