基于生成对抗网络的超声图像超分辨率重建  

Super-resolution reconstruction of ultrasound images based on a generative adversarial network

作　　者：唐真迪 何连海 彭博[1,2,3] 谢盛华[2,3] TANG Zhendi;HE Lianhai;PENG Bo;XIE Shenghua(School of Computer Science,Southwest Petroleum University,Chengdu Sichuan 610500,China;Cardiovascular Ultrasound&Non-invasive Cardiology Department,Sichuan Provincial People’s Hospital,Chengdu Sichuan 610072,China;Key Laboratory of Ultrasound in Cardiac Electrophysiology Biomechanics of Sichuan,Sichuan Provincial People’s Hospital,Chengdu Sichuan 610072,China)

机构地区：[1]西南石油大学计算机科学学院,四川成都610500 [2]四川省人民医院心血管超声及心功能科,四川成都610072 [3]四川省人民医院超声心脏电生理学与生物力学四川省重点实验室,四川成都610072

出　　处：《太赫兹科学与电子信息学报》2023年第5期677-683,共7页Journal of Terahertz Science and Electronic Information Technology

基　　金：四川省科技厅应用基础研究资助项目(2021YJ0248);四川省科技厅应用基础研究资助项目(2018JY0649);四川省成都市科技局国际合作资助项目(2019-GH02-00040-HZ)。

摘　　要：针对医学超声图像的分辨率低而导致视觉效果差的问题,使用基于神经网络的图像超分辨率(SR)重建方法提升医学超声图像的分辨率。采用针对自然图像超分辨率重建的生成对抗网络(SRGAN)作为基本方法,通过减少2个输入通道和删除1个残差块对该网络的结构进行更改,并且改进网络损失函数,新增模糊处理数据集,使该网络适应医学超声图像所具备的灰度图像、散斑纹理单一等特点,从而重建出放大4倍的边缘清晰没有伪影的医学超声图像。将改进SRGAN与原始SRGAN的结果相比,峰值信噪比(PSNR)和结构相似性(SSIM)分别有1.792 dB和3.907%的提升;与传统双立方插值的结果相比,PSNR和SSIM分别有2.172 dB和8.732%的提升。To tackle with the problem of poor visual effects caused by low-resolution of medical ultrasound images,a neural network based image Super-Resolution(SR)reconstruction approach is employed to improve the resolution of medical ultrasound images.Based on the Generative Adversarial Network for Super-Resolution(SRGAN),the structure of the network is changed by reducing two input channels and deleting a residual block.A fuzzy dataset is added and the loss function of the network is improved according to the characteristics of medical ultrasound images,such as gray-scale image and single speckle texture,so that the network is adapted to reconstruct the clear edges with 4 times magnification of medical ultrasound images without artifacts.Comparing the results of the improved SRGAN with the original SRGAN,the Peak Signal-to-Noise Ratio(PSNR)and Structural SIMilarity(SSIM)are increased by 1.792 dB and 3.907%respectively;compared with Bicubic interpolation,the PSNR and SSIM are increased by 2.172%dB and 8.732%respectively.

关 键 词：超分辨率重建 生成对抗网络 乳腺超声图像 残差块 亚像素卷积层 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]