摘要
目的 探讨基于深度学习中的多激活卷积神经网络(MACNN)自动学习模型(简称MACNN模型)对膝关节软骨损伤的诊断效能。方法 以大样本膝关节软骨MRI图像为基础组成MACNN模型训练集,通过卷积、池化、全连接激活及分类输出等步骤对图像信息逐层交互分析,自动提取表征并学习隐藏关系,完成MACNN模型训练。选择健康志愿者23例,膝关节软骨损伤患者117例,其中膝关节镜诊断分级Ⅰ级21例、Ⅱ级39例、Ⅲ级30例、Ⅳ级27例。收集膝关节软骨损伤患者术前及健康志愿者入组后的MRI图像共2035幅,以此为基础分析传统人工阅片、支持向量机(SVM)分类及MACNN模型三种方法对患者膝关节软骨损伤的总诊断效能及分级诊断效能。结果 MACNN模型诊断膝关节软骨损伤的AUC高于传统人工阅片、SVM分类,特异性和准确性均高于传统人工阅片、SVM分类,敏感性高于传统人工阅片(P<0.05或<0.01)。MACNN模型诊断Ⅰ~Ⅳ级膝关节软骨损伤的AUC均高于传统人工阅片、SVM分类;诊断Ⅰ级膝关节软骨损伤的特异性高于传统人工阅片、SVM分类,准确性高于传统人工阅片(P<0.05或<0.01);诊断Ⅱ级膝关节软骨损伤的敏感性、特异性及准确性均高于传统人工阅片、SVM分类(P<0.05或<0.01);诊断Ⅲ级膝关节软骨损伤的敏感性、特异性及准确性均高于传统人工阅片(P均<0.01);诊断Ⅳ级膝关节软骨损伤的特异性高于传统人工阅片(P<0.05)。结论 MACNN模型对膝关节软骨损伤患者的总诊断效能及分级诊断效能均较高,尤其在Ⅰ级及Ⅱ级膝关节软骨损伤诊断中有明显优势。
Objective To explore the classification efficiency of the different knee cartilage damage levels based on the automatic learning model designed by multi-active convolutional neural network(MACNN)method.Methods MACNN model training set was formed based on MRI images of knee cartilage with large samples,through convolution,pooling,full connection activation,classification output and other steps,layer by layer interactive analysis of image information was conducted,and automatic extraction was performed and hidden relations were learned,so as to train the prediction ability of MACNN models.A total of 2035 MRI images of normal knee cartilage and different grades injuried knee cartilage were selected(23 normal subjects,117 knee cartilage damage subjects including gradeⅠof 21 cases,gradeⅡof 39 cases,gradeⅢof 30 cases and gradeⅣof 27 cases confirmed by the arthroscopic surgery).Then we compared the diagnostic efficiency of traditional manual diagnosis,traditional machine learning classification(SVM classification)and MACNN model to study whether the MACNN model is the best way to classify the knee cartilage damage.Results The MACNN model had the highest AUC for the knee cartilage damage diagnosis,the specificity and accuracy were higher than those of the traditional manual reading and SVM classification,and the sensitivity was higher than that of the traditional manual reading(P<0.05 or P<0.01);the MACNN model had the highest AUC in the gradeⅠ-Ⅳof knee cartilage damage diagnosis;the specificity of diagnosis was higher than that of the traditional manual reading and SVM classification in the gradeⅠ,and the accuracy of diagnosis was higher than that of the traditional manual reading(P<0.05 or P<0.01);the specificity,sensitivity,and accuracy of diagnosis were higher than those of the traditional manual reading and SVM classification in the gradeⅡ(P<0.05 or P<0.01);the specificity,sensitivity,and accuracy of diagnosis were higher than those of the traditional manual reading in the gradeⅢ(P<0.01);the specificity of diagnosis was higher than that of the traditional manual reading in the gradeⅣ(P<0.05).Conclusion The MACNN model has higher overall diagnostic efficacy and classification diagnostic efficacy for patients with knee cartilage injury,especially in the diagnosis of gradeⅠandⅡknee cartilage injuries.
作者
杨贵昌
钟妍其
朱彦
刘哲
王冬青
YANG Guichang;ZHONG Yanqi;ZHU Yan;LIU Zhe;WANG Dongqing(The First Hospital of Zibo,Zibo 255200,China;不详)
出处
《山东医药》
CAS
2019年第35期29-32,共4页
Shandong Medical Journal
基金
国家自然科学基金面上项目(61772242)
关键词
膝关节软骨损伤
多激活卷积神经网络
磁共振成像
自动学习模型
knee cartilage injury
multi-activate convolutional neural network
magnetic resonance imaging
automatic learning model
作者简介
第一作者简介:杨贵昌(1967-),男,副主任医师,研究方向为疾病相关磁共振诊断。E-mail:hpwangyang@126.com;通信作者:朱彦(1984-),男,主治医师,研究方向为疾病相关磁共振诊断。E-mail:salary_hi@126.com。
