Direct Diffusion Bridge for High MB-count Image Generation in Ultrasound Localization Microscopy

Ultrasound localization microscopy (ULM) overcomes the acoustic diffraction limit, while maintaining imaging penetration, offering subwavelength resolution for microvascular visualization. However, reconstructing high-quality ULM images necessitates tracking microbubbles (MBs) across thousands of frames, resulting in extended acquisition times and processing time. Generating high MB-count ULM (HC-ULM) images from low MB-count ULM (LC-ULM) images is expected to decrease the acquisition and processing time. In recent years, diffusion models have gained prominence in the field of natural images generation and are superior to generative adversarial networks (GANs). Generative diffusion models rely on progressive disturbing data to noise, and learning the image distribution to reverse the process from noise to high-fidelity data. Diffusion models learn the high-dimensional distribution of training images by constraining consistency differences to generate images. However, MB images cannot establish robust consistency constraints for image synthesis because of the sparsity. In this work, we introduced Direct Diffusion Bridge Model (DDBM), a novel conditional diffusion model using the Schrödinger Bridge, to construct a diffusion bridge between the distribution of LC-ULM image and the distribution of HC-ULM image, so that the model can generate HC-ULM image from LC-ULM image. DDBM is trained using in vivo rat liver data and the results indicate that the proposed method can generate HC-ULM images that are comparable with true HC-ULM images. The proposed method is highly effective in performing ULM image generation tasks and is potential to promote the clinical application of ULM.