A Fully Convolutional Neural Network for Beamforming Ultrasound Images

Plane wave ultrasound imaging is one of the fastest ultrasound methods available to reduce latency for ultrasound-based robotic tracking tasks. However, the presence of acoustic clutter and speckle in the images can confuse robotic tracking algorithms. In addition, multiple plane wave insonification angles are often necessary to generate good-quality images, which further reduces the speed of the tracking process. To overcome these challenges, we are exploring deep learning as a method to extract pertinent information directly from raw radiofrequency channel data to locate targets of interest from a single plane wave insonification. Particularly, in this work, we trained a deep convolutional neural network (CNN) with 50,000 Field-II simulations corresponding to a single cyst in tissue insonified by one plane wave transmitted at 0 degrees. The simulated cyst radius, axial and lateral positions were varied, along with the simulated tissue sound speed. The output of the training process is an interpretable segmentation mask that is free of clutter and speckle, which we call a CNN-Based image. An additional dataset of 100 simulations was created and two cyst targets in an experimental phantom were imaged to test our approach. The Dice Similarity Coefficient (DSC), representing the overlap between the true cyst location and the cyst location in the CNN-Based image, was 0.91 for simulated data and 0.74 for experimental data. The network was generally sensitive to cyst radius, with mean DSCs increasing from 0.91 to 0.97 when the cyst radius was ≥ 5 mm. A robot controlled ultrasound probe enabled volumetric reconstruction of CNN-Based images, revealing the three-dimensional structure of the two cysts in the phantom. These results demonstrate that a deep neural network trained exclusively with simulated data can generalize to experimental data, which is promising for the development of deep learning methods as an alternative to traditional ultrasound beamforming for robotic tracking tasks.