Challenge on Ultrasound Beamforming with Deep Learning (CUBDL)

Recent developments in deep learning have created immense potential for improving ultrasound beamforming. We organized a Challenge on Ultrasound Beamforming with Deep Learning (CUBDL) to benchmark methodologies in this space with two transmission data types: plane wave and focused transmissions. Plane wave ultrasound transmissions have created new opportunities for ultrafast ultrasound imaging, while focused ultrasound transmissions are more traditional and are widely used in most clinical ultrasound systems available today. For both transmission types, we challenged participants to obtain the best image quality under the fastest possible frame rates. CUBDL organizers solicited datasets from several leading ultrasound groups around the world and received a total of 106 data sequences including in vivo, ex vivo, simulated, and experimental phantom datasets. These submissions formed our test datasets, which were not released to participants while the challenge was open. The challenge was composed of three optional tasks (one including two subtasks) that were evaluated using the test datasets. Participants had the option to provide their results for a minimum of one up to a maximum of four tasks or subtasks: (1) beamforming with deep learning after a single plane wave transmission, which had two subtasks to either (a) match or (b) exceed traditional image quality metrics obtained with multiple plane wave transmissions; (2) beamforming with deep learning after a few plane wave transmissions; (3) beamforming with deep learning to achieve dynamic transmit focusing from datasets acquired with a single transmit focus. Evaluation included image quality metrics as well as network complexity metrics. A challenge website was created to provide information and updates: https://cubdl.jhu.edu/.