Efficient Ultrasound Image Enhancement Using Lightweight CNNs




Anjidani, Farid

Journal Title

Journal ISSN

Volume Title



Plane-wave ultrasound imaging allows for very high frame rates. During image reconstruction, conventional delay-and-sum beamforming can be replaced by the quicker Fourier-domain remapping method. Typically, after Fourier-domain reconstruction, postbeamforming interpolation is needed to increase the image grid resolution in the lateral dimension. To achieve this, we propose to use a fast lightweight superresolution convolutional neural network (CNN) operating on the Fourier-beamformed envelope data. Specifically, we train different configurations of well-known Efficient Sub-Pixel Convolutional Neural Network (ESPCN) to perform both 1D and 2D upscaling. First, we pretrain a network using the diverse (non-ultrasound) dataset DIV2K. Then, we apply transfer learning on a small augmented dataset of public-domain experimental ultrasound images. Our results demonstrate that our approach is capable of producing enhanced ultrasound images having higher quality compared to non-CNN interpolation options and conventional delay-and-sum beamforming.



Bicubic Interpolation, Coherent Plane-Wave Compounding, Convolutional Neural Network, Delay-and-Sum (Beamforming), Fast Fourier Transform, Fourier Interpolation, Full Width at Half Maximum, Mean Squared Logarithmic Error, Plane Wave Imaging, Temme-Mueller (Migration), UltraSound, Single-Image Super-Resolution, beamforming, DAS, superresolution