Ultrasound brain imaging with latent diffusion models

Recent advancements in medical imaging have demonstrated the potential of full-waveform inversion (FWI) of ultrasound data, a technique developed in geophysics, to produce high-fidelity images of the human body at remarkable resolutions. For the first time, FWI has enabled the visualisation of fine anatomical structures in the human brain using ultrasound tomography, marking an unprecedented achievement in healthcare research. However, FWI’s ill-posed nature, coupled with its computational demands, renders it impractical for immediate clinical deployment. Deep neural networks, meanwhile, have driven progress across scientific domains, with diffusion models gaining recognition for their ability to model complex probability distributions and synthesise high-quality data. In this work, I propose a method for recovering high-fidelity, high-resolution images of the human brain with conditioned latent diffusion models (LDMs), as a means to bypass the limitations of FWI. The research unfolds in three stages, each offering distinct contributions that advance ultrasound brain imaging. The first introduces a novel loss function, the Wiener Loss, which employs convolutional analysis to compare data without relying on local, element-wise relationships or costly model inferences. This enables high-fidelity analysis of complex structures, such as the human brain, making it well-suited to deep learning tasks requiring contextual and perceptual accuracy. The second stage constructs and trains a condition-free LDM on a limited dataset of acoustic neuroimages, demonstrating its ability to rapidly generate statistically conforming, diverse synthetic samples for data augmentation. The third employs conditional LDMs to learn the statistical mapping between acoustic velocity models of the brain and associated ultrasound data in a synthetic setting, enabling high-fidelity image recovery at near-real time. Although time and resource constraints limited success in the final stage, preliminary results show strong potential for this approach as a fast, efficient alternative to FWI. Its independence from physical constraints suggests broader applicability to other data inversion challenges.