XGeM is a 6.7 billion-parameter multimodal generative model proposed to address the challenges of AI utilization in medical imaging: data scarcity, privacy concerns, and the need for robust multimodal integration. It constructs a shared latent space via contrastive learning and introduces a novel multi-prompt training strategy that can condition on an arbitrary subset of input modalities, thereby supporting synthesis across various arbitrary modalities. We compare it with competing models using the MIMIC-CXR dataset, and evaluate the realism and clinical relevance of the generated data through a visual Turing test targeting expert radiologists. We demonstrate that it can be utilized to address healthcare data challenges such as data anonymization, class imbalance, and data scarcity.
