3D Ultrafast Ultrasound Image Quality Enhancement using 3D Deep Convolutional Neural Networks

Three-dimensional (3D) ultrafast ultrasound imaging using matrix-array transducers and unfocused wave emission has huge potential for clinical diagnosis. However, the pursuit of high frame rates often compromises image quality, and achieving satisfactory 3D volume quality typically demands advanced transducers and hardware systems. These factors have significantly hindered the widespread adoption of 3D ultrafast ultrasound. Recent researches have explored deep learning techniques to enhance conventional 2D ultrasound images, however, to our knowledge, none have proposed using a 3D convolutional neural network (3D-CNN) to optimize 3D ultrasound volumes. Thus, a 3D ultrafast ultrasound image quality enhancement method using a 3D U-Net trained using simulated and phantom low-high quality volumes pairs is proposed in this work. Lateral resolution and contrast are calculated to evaluate the performance of the model. The results show that the proposed method could produce high quality volumes equivalent to the compounding of 14 plane waves in terms of both contrast and lateral resolution.