Machine Learning-Based B-mode Imaging
Ultrasound B-mode images display the “echogenicity”, or acoustic brightness, of the underlying soft tissue. Traditional B-mode images are formed using delay-and-sum (DAS) beamforming, which displays the magnitude of the mean complex channel signal. DAS is a robust technique that is used ubiquitously in medical ultrasound imaging systems, but is fundamentally subject to noise artifacts such as speckle. Machine learning and deep learning techniques present an alternative way to tackle echogenicity estimation. Rather than using a deterministic algorithm such as DAS, neural networks are trained empirically to reconstruct optimal B-mode images as quantified by metrics such as mean absolute error, mean squared error, structural similarity, etc. We have recently demonstrated that simple fully convolutional neural networks produce more accurate echogenicity estimates than DAS in ultrasound simulations.
An open source implementation of neural network-based B-mode imaging is provided here:
Training, validation, and testing data are also provided:
This code is free to use and is covered by the Apache v2 license. Please cite the following reference when using this code or data:
D. Hyun, L. L. Brickson, K. T. Looby, and J. J. Dahl. "Beamforming and Speckle Reduction Using Neural Networks." IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, 66(5), pp. 898-910, May 2019. doi: 10.1109/TUFFC.2019.2903795