Temporally Consistent and Controllable Video Generation of 2D Cine CMR via Latent Space Motion Modeling

The article presents a novel generative method for synthesizing temporally coherent and anatomically consistent cine cardiac magnetic resonance (CMR) sequences using a text-to-video framework. The approach employs a fine-tuned diffusion model to create initial frames based on clinical text prompts, followed by a latent flow model that generates complete cardiac motion while maintaining spatial consistency, achieving a Fréchet Inception Distance (FID) of 31.68 and a CLIP score of 31.04. This method addresses data scarcity in medical imaging by enabling the generation of high-fidelity, diverse cardiac sequences, which could significantly enhance training datasets for AI models in healthcare.