FeatuReBeam: Efficient Beamforming for High-Frame Ultrasound Imaging Using Adaptive Feature Reuse With <i>(m,k)</i>-Guarantee

Existing tissue adaptive deep learning-based ultrasound image reconstruction models are computationally intensive, and achieving high frame rates particularly in the application of B-mode imaging on resource-constrained devices is challenging. Based on a traditional firm guarantee real-time model, i.e., the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">(m,k)</i>-model, we propose a methodology that can take advantage of the similarity among feature maps generated by continuous frames to reduce the redundant computations and, thus, significantly improve the throughput and energy efficiency. A comprehensive evaluation on <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">in-vitro</i> and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">in-vivo</i> data using U-Net and GoogLeNet-based beamformers demonstrated over a 50% reduction in computational complexity while preserving the original image quality. The implementation of the proposed methodology on the Xilinx ZU7EV FPGA achieved latency improvement of over 45% and 46% for the U-Net and GoogLeNet models, respectively. Additionally, when compared to state-of-the-art filter pruning approaches, including magnitude-based and feature map-based filter pruning, the proposed methodology outperformed both in terms of computational reduction and output image quality, achieving over a 14% improvement on the Xilinx ZU7EV FPGA.