文章摘要

文章信息/Info

Title:: High-quality far-focused ultrasound image reconstruction based on parallel attention mechanism and deep neural network

Author(s):: LI Yuantao; XIE Zhongwen; PENG Hu; Department of Biomedical Engineering, Hefei University of Technology, Hefei 230009, China

Keywords:: Ultrasound imaging; High-quality image reconstruction; Deep learning; Dual attention

摘要:: 获得高质量的超声图像,通常需要做多次超声波的聚焦发射，为提高成像帧率而减少发射次数会降低成像质量。为解决该问题，本研究基于深度学习算法，使用Unet辅以并行注意力机制重建高质量远聚焦波超声图像。该方法最少能以16次聚焦发射的回波数据重建出原128次发射才能得到的高质量图像。本研究结合超声数据的特点，使用三维数据体作为输入，而非单张超声图像到另一张图像的直接映射重建。该方法在计算机仿真、物理体模以及人体实验上都得到了很好的结果。实验结果证明该方法在不牺牲图像质量的前提下，有效地降低了数据传输速率。

Abstract:: To obtain high-quality ultrasound images usually requires multiple focused ultrasound transmissions.Reducing the number of transmissions to increase the imaging frame rate will influence the imaging quality. To solve this problem, we used Unet supplemented by a parallel attention mechanism based on deep learning algorithm to reconstruct high-quality far-focused ultrasound image. It could reconstruct the high-quality images that could only be obtained by the original 128 shots from the echo data of 16 focused shots at least. We combined the characteristics of ultrasound data and used three-dimensional data volume as input instead of direct mapping and reconstructing from a single ultrasound image to another image. it obtained good results in computer simulation, physical phantom and human experiment. Experimental results demonstrate that this method effectively reduces the data rate without sacrificing the image quality.