文献类型：期刊文献

中文题名：基于改进双路径网络的上肢肌肉骨骼异常检测

英文题名：Detection of upper limb musculoskeletal abnormality based on improved dual path network

作者：黄彩云[1];陈德武[2];何吉福[1];胡艺[1];王楠[1];陈沛[1]

第一作者：黄彩云

机构：[1]甘肃中医药大学体育健康学院,甘肃兰州730000;[2]中国石油勘探开发研究院西北分院地球物理研究所,甘肃兰州730020

第一机构：甘肃中医药大学体育健康学院

年份：2022

卷号：52

期号：3

起止页码：25

中文期刊名：山东大学学报（工学版）

外文期刊名：Journal of Shandong University（Engineering Science）

收录：CSTPCD;;北大核心:【北大核心2020】；CSCD:【CSCD_E2021_2022】；

基金：甘肃省自然科学基金资助项目(1606RJZA197);甘肃省教育科学“十四五”规划2021年度“双减”专项课题(GS[2021]GHBZX269);甘肃中医药大学科技创新项目(30740301)。

语种：中文

中文关键词：医学图像病变检测;上肢肌肉骨骼异常;双路径网络;深度残差网络;密集连接卷积网络

外文关键词：medical image lesion detection;upper limb musculoskeletal abnormality;dual path network;deep residual network;densely connected convolutional network

摘要：针对传统的基于专家知识经验医学图像病变检测算法存在稳定性较差、计算复杂度较高、无法适应新出现病例等问题,提出一种基于改进双路径网络(dual path networks with skip connections,SkipConn_DPN)的上肢肌肉骨骼X射线照片异常检测方法。借助深度学习卷积神经网络在图像处理领域的优异性能,整合改进的深度残差网络在图像特征重用和密集连接卷积网络可以不断挖掘新图像特征的优点,并引入可以实现不同尺度图像特征融合的多条跳跃连接。在不增加计算资源消耗的前提下,该方法训练的8个SkipConn_DPN-92网络模型对MURA验证集中不同上肢研究类型的准确率均高于相同参数设置下训练的DPN-92、ResNet-101、ResNeXt-101(32×4d)和DenseNet-169网络模型,分别高约1.64%、2.21%、1.87%和3.22%,并且在临床上可以实现上肢肌肉骨骼异常的实时检测。提出的方法易于实现软件模块,可以作为放射科医生初步诊断的可视化辅助工具,具有良好的应用前景。
Aiming at the problems of poor stability,high computational complexity,and inability to adapt to emerging cases in traditional medical image lesion detection algorithms based on expert knowledge and experience,an anomaly detection method based on an improved dual path network(SkipConn_DPN)for upper extremity musculoskeletal X-ray photographs was proposed.This method integrated the advantages of the improved deep residual network in image feature reuse and densely connected convolutional network in mining new image features continuously,and introduced multiple skip connections that could realize the fusion of image features at different scales.Without increasing computational resource consumption,the accuracy of the 8 trained SkipConn_DPN-92 network models on different upper limb research types in the MURA validation set was about 1.64%,2.14%,1.8%and 3.15%higher on average than that of the DPN-92,ResNet-101,ResNeXt-101(32×4d)and DenseNet-169 network models,which were trained under the same parameter settings,and the real-time detection of upper limb musculoskeletal abnormalities could be clinically realized.The proposed method was easy to implement software module,could be used as a visual aid for radiologists to make preliminary diagnosis,and had good application prospects.