Synthesis of realistic medical images with pathologies using diffusion models with application to lung CT and mammography

The scarcity of annotated medical images poses a significant challenge to developing and training accurate AI models for the detection of tumors and other pathologies. To address this, we introduce a novel method for synthesizing realistic medical images using Multi-Conditioned Denoising Diffusion Probabilistic Models, capable of generating images with and without tumors. Our approach leverages existing datasets to train the diffusion models, ensuring the synthetic images closely mimic real-world medical images. This controlled synthesis process allows for the creation of diverse datasets, enhancing the variability and richness of training data. By generating high-quality synthetic medical images, we aim to overcome the limitations of data scarcity and improve the performance and generalizability of AI models. While our primary goal is to increase the volume and diversity of training data, this method also holds potential benefits for underrepresented population groups by facilitating the inclusion of more varied demographic and pathological characteristics. Our results indicate that the diffusion model-generated medical images are indistinguishable from real images by radiologists, demonstrating their potential for effective use in AI model training. We also found that enriching sparse training data with our synthetic images can improve the accuracy of pathology detection AI classifiers. This innovative approach promises to significantly advance the field of AI-driven medical imaging, leading to more accurate and reliable diagnostic tools. Our paper presents first results on two specific important applications, lung CT and mammography.