基于视频多帧融合的医学超声图像超分辨率重建方法  

Medical Ultrasound Image Super-resolution Reconstruction Based on Video Multi-frame Fusion

作　　者：赵冉 袁家斌[1] 范利利 ZHAO Ran;YUAN Jiabin;FAN Lili(School of Computer Science and Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China)

机构地区：[1]南京航空航天大学计算机科学与技术学院,南京211106

出　　处：《计算机科学》2023年第7期143-151,共9页Computer Science

基　　金：国家重点研发计划(2017YFB0802303);国家自然科学基金(62076127)。

摘　　要：医学超声成像是临床诊断中应用最广泛的成像方式之一。目前超声图像普遍存在分辨率和对比度较低的问题,并且成像过程易受噪声污染。图像超分辨率重建技术被广泛用于改善超声图像的质量。然而,已有的研究工作缺乏对超声视频帧之间互补信息的充分利用,因此效果并不理想。针对此问题,提出了一种基于视频多帧融合的医学超声图像超分辨率重建方法。首先,构建了一个基于卷积神经网络的无监督多帧融合模型,该模型通过对连续的多帧图像进行特征融合,得到具有丰富信息的融合特征图像;然后,建立一个基于无数据知识蒸馏的轻量级图像超分辨率重建模型,通过训练融合特征图像得到教师超分辨率网络,利用训练好的教师网络和生成对抗网络获取的训练数据得到轻量级学生超分网络,最终得到高质量的医学超声图像;最后,在大型超声数据集上进行实验,采用两种图像客观评价指标以及图像分类任务进行评估。结果表明,所提方法与8种已有的图像超分辨率重建方法相比,在提高超声图像分辨率的同时,获得了包含更多信息且具有更高对比度的超声图像。此外,所提方法得到的超分辨率图像在分类网络的识别准确率可达到97.30%,明显优于其他方法,可提高临床诊断效率与准确性。Medical ultrasound imaging is one of the most widely used methods in clinical diagnosis.However,the resolution and contrast of ultrasound images are low,and the noise is serious.Image Super-resolution is widely used to improve the quality of ultrasound images.However,the exsisting studies lack full use of complementary information in ultrasound video frames,so the results are not ideal.To solve this problem,this paper proposes a super-resolution method of medical ultrasound images based on video multi-frame fusion.Firstly,it builds a multi-frame fusion model based on convolutional neural network for ultrasound images.The model obtains the fused image with rich information by feature fusion of continuous multi-frame images.Then,it builds a lightweight image super-resolution model based on data-free knowledge distillation.The model trains fused feature images to obtain a teacher super-resolution model.To obtain the final high-quality medical ultrasound image,it creates a lightweight student super-resolution model using the trained teacher model and training data from the generative adversarial network.Finally,the proposed method is tested on large ultrasound datasets,using two objective image evaluation indicators and results on image classification to evaluate its performance.The results show that compared with other methods,the proposed method not only improves the resolution of ultrasound images,but also obtains ultrasound images with more information and higher contrast.The recognition accuracy of the super-resolution image obtained by this method in the classification network can reach 97.30%,which is much higher than that of previous approaches and can increase productivity and the precision of clinical diagnoses.

关 键 词：医学超声图像 多帧融合 超分辨率重建 无数据知识蒸馏 卷积神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]