Improving Image Quality of Single Plane Wave Ultrasound via Deep Learning Based Channel Compounding

Improving Image Quality of Single Plane Wave Ultrasound via Deep Learning Based Channel Compounding Sven Rothlübbers1, Hannah Strohm1, Klaus Eickel2, Jürgen Jenne1, Vincent Kuhlen1, David Sinden1, Matthias Günther1,2, (1) Fraunhofer MEVIS, Bremen, Germany, (2) University of Bremen, Bremen, Germany Background, Motivation and Objective Planewave imaging offers high frame rates. However, when using only a few plane waves per image, the quality of the result deteriorates. Reconstruction methods exist which are suitable to create high quality images, but require long computation times. There is a clear need for fast reconstruction algorithms which operates on only few acquisitions, addressed here as a contribution to task 1 of the CUBDL-challenge [1]. Statement of Contribution/Methods This work presents an architecture that computes weighting factors for individual pixel intensities obtained from unweighted channel summation. The network is trained to reproduce a high-quality target image obtained from modified multiangle united sign coherence factor (USCF) [2] reconstruction of phantom raw data (CIRS structural phantom, DiPhAS US system). We use the time delayed, magnitude normalized, but not yet channel-summed complex data after beamforming as input. The network consists of four layers performing 1D convolutions with ReLu activations along the complex valued ultrasound channel axis. Kernel sizes 65,15,15,3 and feature map sizes 8,8,8,1 are used. Finally, the ultrasound channels are summed to form the pixel weights. The loss is computed as a linear combination of MSE and MSSSIM loss on the log compressed, normalized images. Results/Discussion From a single acquisition, the trained network is able to produce visually more appealing images than single shot SCF-beamformed [3] reference images (see Figure 1). There are a number of advantages of the architecture, firstly, in contrast to fully-connected networks [4], the proposed fully-convolutional architecture allows for both the handling of images of varying sizes and ultrasound channels. Furthermore, as the network performs the reconstruction after applying time-delays, it can be integrated into other reconstruction pipelines. With fewer layers, the architecture has a lower network complexity than existing deep learning models [5]. Acknowledgements We are grateful for Fraunhofer Funding Discover 600725 UltraDeep and Prepare 601100 Theranus. Figure 1: Comparison of network output to multi and single angle reference reconstructions on unseen test data from PICMUS [6] dataset. [1] Challenge on Ultrasound Beamforming with Deep Learning (CUBDL), IEEE Dataport, 2019. [Online]. Available: http://dx.doi.org/10.21227/f0hn-8f92. Accessed: Jun. 22, (2020); [2] C. Yang, Y. Jiao, J. Jiang, T. Jiang, X. Yiwen, C. Yaoyao “A United Sign Coherence Factor Beamformer for Coherent Plane-Wave Compounding with Improved Contrast” Appl. Sci. 10(7):2250. doi:10.3390/app10072250 (2020); [3] J. Camacho, M. Parrilla, C. Fritsch “Phase Coherence Imaging”, IEEE Trans Ultrason Ferroelectr Freq Control. 56(5):958-974 doi:10.1109/TUFFC.2009.1128 (2009); [4] B. Luijten et al. Deep Learning for Fast Adaptive Beamforming, ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Brighton, United Kingdom, pp. 1333-1337, doi: 10.1109/ICASSP.2019.8683478 (2019); [5] S. Khan and J. Huh and J. C. Ye “Adaptive and Compressive Beamforming Using Deep Learning for Medical Ultrasound”, https://arxiv.org/abs/1907.10257 (2019); [6]H. Liebgott, A. Rodriguez-Molares, F. Cervenansky, J. A. Jensen and O. Bernard, Plane-Wave Imaging Challenge in Medical Ultrasound, 2016 IEEE International Ultrasonics Symposium (IUS), Tours, 2016, pp. 1-4, doi: 10.1109/ULTSYM.2016.7728908. (2016).