融合多尺度图像的密集神经网络肺部肿瘤识别算法  

Lung tumor image recognition algorithm with densenet fusion multi-scale images

作　　者：周涛 霍兵强 陆惠玲[4] 马宗军[5] 叶鑫宇 董雅丽 刘珊 ZHOU Tao;HUO Bing-qiang;LU Hui-ling;MA Zong-jun;YE Xin-yu;Dong Ya-Li;Liu Shan(School of Computer Science and Engineering,North Minzu University,Yinchuan 750021,China;Key Laboratory of Images&Graphics Intelligent Processing of State Ethnic Affairs Commission,North Minzu University,Yinchuan 750021,China;Ningxia Key Laboratory of Intelligent Information and Big Data Processing,Yinchuan 750021,China;School of Science,Ningxia Medical University,Yinchuan,NingXia 750004,China;Department of Orthopedics,General Hospital of Ningxia Medical University,Yinchuan 750021,China)

机构地区：[1]北方民族大学计算机科学与工程学院,银川宁夏750021 [2]北方民族大学图像图形智能处理国家民委重点实验室,银川宁夏750021 [3]宁夏智能信息与大数据处理重点实验室,银川宁夏750021 [4]宁夏医科大学理学院,银川宁夏750004 [5]宁夏医科大学总医院骨科,银川宁夏750004

出　　处：《光学精密工程》2021年第7期1695-1708,共14页Optics and Precision Engineering

基　　金：国家自然科学基金(No.62062003);宁夏重点研发项目(引才专项)(No.2020BEB04022);北方民族大学引进人才科研启动项目(No.2020KYQD08)。

摘　　要：针对CT模态医学图像采用卷积神经网络训练时的特征提取不充分、特征维度较高等问题,本文提出了基于融合多尺度图像的非负稀疏协同表示分类的密集神经网络肺部肿瘤(Multi Scale DenseNet-NSCR)的识别方法。第一,使用迁移学习将预训练密集神经网络模型初始化参数;第二,将肺部图像预处理,提取多尺度病灶ROI区域;第三,采用多尺度CT图像训练密集神经网络,提取全连接层的特征向量;第四,针对融合特征维度较高问题,采用非负稀疏协同表示分类器(NSCR)对特征向量进行表示,求解系数矩阵;第五,利用残差相似度进行分类。最后,采用AlexNet,DenseNetNet-201模型及三种分类算法(SVM、SRC、NSCR)两两组合模型进行对比试验,实验结果表明,Multiscale-DenseNet-NSCR分类效果优于其它模型,且特异性和灵敏度等各项评价指标也较高,该方法具有较好的鲁棒性和泛化能力。To address the problem of inadequate feature extraction and high feature dimension when CT modal medical images are trained with convolutional neural networks,this paper proposes a method for lung tumor identification using Multi-scale DenseNet-NSCR based on non-negative sparse collaborative representation classification by fusing multi-scale images.First,the parameters of the pre-trained dense neural network model are initialized using migration learning;the lung images are then pre-processed to ex⁃tract multi-scale lesion ROI.Subsequently,the DenseNet is trained using a multi-scale CT dataset to ex⁃tract feature vectors at the full connection layer.To address the problem of the high dimensionality of the fused features,a non-negative,sparse,and collaborative representation(NSCR)classifier is used to rep⁃resent the feature vector and solve the coefficient matrix;the residual similarity is then used for classifica⁃tion.Finally,a comparison test is conducted with the AlexNet,DenseNetNetNet-201 model,and a com⁃bination model of three classification algorithms(SVM,SRC,NSCR).The experimental results show that Multiscale-DenseNet-NSCR classification is better than other models;all evaluation indexes such as specificity and sensitivity are higher,and the method has better robustness and generalization ability.

关 键 词：密集神经网络 多尺度医学图像 迁移学习 NSCR算法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]