Self-supervised ultrasound video frame interpolation based on conditional latent diffusion model

High temporal resolution in ultrasound (US) videos is essential for imaging fast-moving structures, capturing dynamic changes, and accurately detecting diseases. Video frame interpolation (VFI) enhances temporal resolution by generating intermediate frames. Despite advancements in natural VFI, motion-estimation-based deep learning methods struggle with US videos due to their complex motion and image quality. In this work, we introduce the conditional latent diffusion model (LDM) for 2D US VFI for the first time, formulating frame prediction as a conditional generative task. We evaluate our method against state-of-the-art motion-estimation-based approaches using one open-source and one customized ultrasound video dataset. Both qualitative and quantitative analyses demonstrate that the conditional LDM approach noticeably enhances the quality of US VFI by generating sharper and more realistic frames.