Accelerated ultrasound imaging using data-driven dictionary learning and sparsity-penalized interpolation
To reduce data rate for a portable ultrasound imaging system, various compressed sensing approaches have been investigated. However, most of the existing approaches require either hardware changes or accurate forward modeling of wave propagation. To overcome the drawbacks, sparsity-penalized data interpolation algorithms that exploit raw data domain redundancies have been proposed. One of the main limitations of such approaches are, however, extremely high computational complexity due to redundant 3-D transform that captures various edge directions in raw data. To overcome such limitation, this research proposes a computational efficient data-driven non-redundant dictionary learning algorithm. In particular, an interleaved downsampling scheme is proposed, where fully sampled frames are acquired in-between downsampled frames and data-driven dictionaries are learned using the fully sampled data.
we proposed a data-driven dictionary learning method and sparsity-penalized interpolation method to accelerate the ultrasound data acquisition. Compared to the existing method using redundant wave atom transform, the proposed method was two order of magnitude faster and exhibited competitive performance since the resulting transform is non-redundant but still data-adaptive. Experimental results obtained from the carotid confirmed that the proposed method reproduced the missing data effectively and visually equivalent results were obtained compared to the fully sampled data.