Quality Enhancement of Ultrafast Ultrasound Images with Deep Networks and Transfer Learning

Ultrafast ultrasound achieves high frame rates at the expense of image quality. Therefore, the use of convolutional neural networks (CNNs) is currently investigated to mitigate reconstruction artifacts and improve the quality of the images. Promising results are currently achieved by CNNs trained on simulated data, however there is a significant domain gap between training data and in vivo images, impeding the performance of such an approach in practice. In this work, we seek to bridge this domain gap by applying transfer learning (TL) from simulated data to a set of real carotid images. A U-net-based CNN is trained to map radio-frequency images acquired with single plane-wave (PW) acquisitions to target images, obtained from the coherent compounding of 87 PWs. TL is performed by fine-tuning, using a dataset of carotid images, the upsampling arm of the U-net trained on simulated data. Our results demonstrate that the proposed TL training strategy outperforms qualitatively and quantitatively training approaches based solely on simulated or in vivo data.