Fraunhofer MEVIS wins CUBDL Challenge

© Fraunhofer MEVIS
Images reconstructed from single angle plane wave raw data: Reference reconstruction with Sign Coherence Factor (left) and proposed CNN-reconstruction result in which the network had been trained to produce multiangle images reconstructed with United Sign Coherence Factor (right).

Scientists from Fraunhofer MEVIS successfully participated in the Challenge on Ultrasound Beamforming with Deep Learning (CUBDL). The challenge was part of the 2020 IEEE International Ultrasonics Symposium (IUS) that took place in “virtual” Las Vegas from September 7-11.

All CUBDL participants had submitted methods to approach Task 1 of the challenge: “creating a high-quality image from a single plane wave to match a higher quality image created from multiple plane waves”. Among them, Fraunhofer MEVIS’ team around Sven Rothlübbers shares first place with a competitor from Concordia University, Montréal.

While the competing approach could produce better overall image quality, the solution by Fraunhofer MEVIS convinced the jury with its strongly reduced network complexity compared to published solutions. The method applies a few-layer Convolutional Neural Network within the image reconstruction pipeline in order to improve the contrast weighting for individual pixels.

Recent developments in deep learning have created immense potential for ultrasound imaging research. CUBDL participants were challenged to obtain the best image quality under the fastest possible frame rates. The challenge was designed to explore the benefits of using deep learning for a variety of typical imaging setups.