Model-aware deep learning for clutter suppression in contrast-enhanced ultrasounds

Ultrasonography is a low-cost, radiation-free invivo imaging modality. In particular, blood vessel imaging and the measurement of their physical parameters, such as blood flow velocities are of great clinical interest. One prominent solution involves the use of encapsulated gas microbubbles for improved visualization of the vascular bed, which has recently enabled imaging with unprecedented sub-wavelength spatial resolution. A typical preprocessing step to such operations is first separating the microbubble signal from the cluttering tissue signal. Here, we propose two main contributions. The first, is adopting a new model for the acquired contrast enhanced ultrasound signal, which was similarly suggested in the context of ultrasound Doppler processing and dynamic magnetic resonance imaging, and is typically solved via an iterative algorithm. This model leads to a better separation of microbubble signal from the tissue signal than commonly practiced methods. The second, is to utilize the recently proposed technique of learning fast sparse approximations from the field of deep learning to suggest a suitable network architecture for this task with improved convergence speed and accuracy over its iterative counterpart. We compare the performance of the suggested network on both simulations and in-vivo rat brain scans, with a commonly practiced deep-network architecture and the fast iterative shrinkage/thresholding algorithm. We show that our architecture exhibits better image quality and contrast.