DeepCEUS: Interleaved Signals Estimation in Checkerboard Imaging for Contrast Media Imaging using Context-aware Deep Learning

The extraction of harmonic signals for contrast imaging is achieved by using filters and N-pulse imaging techniques (e.g., pulse inversion (N≥2) or checkerboard apertures technique (N≥3)). On one hand, the former is limited by spectral leakage, while on the other hand, the latter suffers from an N-fold reduction in frame rate and increased susceptibility to motion artifacts resulting in incomplete removal of the fundamental echoes. In this work, we propose a context-aware signal-based deep learning approach (DeepCEUS) enabling single shot focused harmonic imaging for contrast media. This is accomplished by implementing an asymmetric convolutional autoencoder, to learn the nonlinear mapping between the backscattered echo from the full-aperture transmitted pulse and the combination of the interleaved pulses used in the conventional checkerboard apertures harmonic imaging technique. With the presented approach, the frame rate of focused contrast imaging can be almost tripled, while maintaining the reconstructed image resolution and contrast. Furthermore, an obvious reduction of noise and motion artefacts can be achieved.