Improving Contrast of Fundamental Ultrasound Imaging Using a Deep Neural Network

Harmonic imaging is used as default in many ultrasound imaging settings due to its advantages in contrast and resolution compared to fundamental imaging. However, the nonlinear harmonic generation process and higher imaging frequencies induce a strong loss of penetration when using harmonic vs fundamental imaging. It would be ideal to have the best of both worlds: a high-contrast mode with high penetration. We propose a deep learning approach that relies on a convolutional neural network (CNN) to improve the contrast of fundamental images. We used detected, log-compressed image data before scan conversion collected from commercial phantoms for training and validation. We captured fundamental and corresponding harmonic images and taught the network to learn the harmonic image from the fundamental one. We observed the expected increase in the contrast using the proposed CNN method with fundamental inputs. The proposed approach is promising for developing dedicated deep-learning networks to improve the real-time ultrasound image formation process.