基于分步目标定位的腰椎间盘自动诊断方法-《中国医学物理学杂志》

文章信息/Info

Title:: Automatic diagnosis of lumbar intervertebral disc herniation based on step-by-step target positioning

作者:: 巩稼民1; 2; 杨红蕊1; 郭庆庆2; 蒋杰伟2; 潘琼3; 马豆豆2; 高燕军4; 1.西安邮电大学通信与信息工程学院，陕西西安 710121； 2.西安邮电大学电子工程学院，陕西西安 710121； 3.西北农林科技大学理学院，陕西西安 712100； 4.西安市第三医院医学影像科，陕西西安 710071

Author(s):: GONG Jiamin1; 2; YANG Hongrui1; GUO Qingqing2; JIANG Jiewei2; PAN Qiong3; MA Doudou2; GAO Yanjun4; 1. School of Communication and Information Engineering, Xian University of Posts and Telecommunications, Xian 710121, China 2. School of Electronic Engineering, Xian University of Posts and Telecommunications, Xian 710121, China 3. School of Science, Northwest Agriculture and Forestry University, Xi an 712100, China 4. Department of Medical Imaging, Xian No.3 Hospital, Xian 710071, China

Keywords:: Keywords: lumbar intervertebral disc herniation step-by-step target positioning Faster R-CNN network improved residual convolutional neural network computer-aided diagnosis system

摘要:: 针对当前腰椎间盘自动诊断方法存在的准确率偏低的问题，提出一种基于分步目标定位的计算机辅助诊断方法。该方法首先使用Faster R-CNN目标定位网络预处理腰椎间盘影像，去除韧带以及周围噪声区域，获得腰椎间盘的轮廓区域；然后放大定位的间盘轮廓3倍，再次利用Faster R-CNN网络精细化定位病灶区域，从而解决因病灶目标太小而无法准确定位的问题；最后，将病灶区域输入到改进的残差卷积神经网络中以提取高层特征和严重性分级，改进的残差卷积神经网络（ResNet-20）通过建立短路机制以提高分类器的准确率。实验结果表明，相较于传统的诊断方法，该方法将腰椎间盘突出的诊断准确率提升5.1%。

Abstract:: Abstract: A computer-aided diagnosis method based on step-by-step target positioning (SSTP) is proposed for solving the problem of low accuracy in current methods for the automatic diagnosis of lumbar intervertebral disc herniation. Firstly, Faster R-CNN target positioning network is used to preprocess lumbar intervertebral disc images, remove ligaments and surrounding noise areas, and obtain the contour of lumbar intervertebral disc. Then, the contour of the located disc is enlarged by 3 times, and Faster R-CNN network is further applied to finely locate the focus area, thus solving the problem of inaccurate positioning due to the small focus. Finally, the focus area is input into the improved residual convolution neural network to extract high-level features and to grade the severity. The improved residual convolutional neural network (ResNet-20) improves the classifier accuracy by establishing a short-circuit mechanism. Experimental results show that the proposed method improves diagnostic accuracy of lumbar intervertebral disc herniation by 5.1% in comparison with traditional diagnostic methods.