Study on Attenuation Tomography Using Low-frequency Ultrasound Differential Imaging with Variational Autoencoder for Human Thorax

Ultrasound (US) is widely employed in medical imaging, but its use in the human thorax is constrained by high attenuation. Recently, researchers have addressed this limitation by utilizing low-frequency US (10 kHz to 750 kHz), which can penetrate the thorax. Studies indicate that low-frequency US can monitor the respiratory system through attenuation factor differential imaging, reconstructing variations in the thorax attenuation factor based on changes in transmitted US energy during respiration. In this paper, we propose a novel inversion method using a neural network structure, the variational autoencoder (VAE), to integrate prior knowledge of thorax structure. The VAE is trained to compress high-dimensional data into a low-dimensional code and subsequently reconstruct it. During inversion, the latent code of the domain of interest (DOI) is first reconstructed, followed by decoding to obtain the high-dimensional attenuation distribution, effectively incorporating structural prior information. Numerical simulations with both training and test set models demonstrate the superiority of the proposed method.