Deep Learning Based Ultrasound Image Reconstruction Method: A Time Coherence Study

Recently, deep learning entered the ultrasound (US) image reconstruction community, demonstrating unprecedented performances on image reconstruction tasks. The use of deep neural networks to reconstruct, restore or enhance US images has been challenged on its capability to preserve time-coherence, mostly because of their inherent non-linear properties. Most novel image reconstruction methods are typically only evaluated on static images, lacking any demonstration of their potential applicability to other imaging modes, such as vector flow imaging (VFI) and shear-wave elastography, which heavily rely on the time-coherence of consecutive reconstructed images. In this work, we demonstrate that our previously proposed convolutional neural network (CNN)-based image restoration approach, trained exclusively to improve the quality of static images, does not harm the time-coherence of consecutive frames in the context of VFI. Using dynamical numerical phantoms inspired by the synthetic aperture vector flow imaging (SA-VFI) challenge, we quantitatively show that the use of such an image restoration technique does not damage vector flow estimations, computed with a state-of-the-art speckle tracking algorithm, on simple phantoms with constant echogenicity, and that it even has the potential to improve such estimations in more complex scenarios.