基于半监督网络的前交叉韧带损伤膝关节磁共振诊断辅助研究  

Diagnosis of anterior cruciate ligament injuries in MRI images based on semi-supervised residual network deep learning

作　　者：危俊杰 张程远[2,3] 姜智瀚 刘坤 孔薇[1] 袁锋[2] WEI Junjie;ZHANG Chengyuan;JIANG Zhihan;LIU Kun;KONG Wei;YUAN Feng(School of Information Engineering,Shanghai Maritime University,Shanghai 201306,China;Department of Orthopedics,Shanghai Sixth People's Hospital Affiliated to Shanghai Jiao Tong University School of Medicine,Shanghai 200233,China;Department of Orthopedics,East Campus of Shanghai Sixth People's Hospital Affiliated to Shanghai College of Medicine and Health Sciences,Shanghai 201306,China)

机构地区：[1]上海海事大学信息工程学院,上海201306 [2]上海交通大学医学院附属第六人民医院骨科,上海200233 [3]上海健康医学院附属第六人民医院东院骨科,上海201306

出　　处：《同济大学学报（医学版）》2024年第4期503-509,共7页Journal of Tongji University(Medical Science)

基　　金：上海市科学技术委员会项目(20Y11913300);上海市申康医院发展中心项目(SHDC12020118);上海市浦东新区卫生健康委员会项目(PW2022D-11);上海健康医学院校级科研项目(SSF-23-14-004)。

摘　　要：目的本研究基于半监督算法残差网络(semi-supervised algorithm Residual network,SMRNet)的深度学习方法,探索其在计算机辅助自主分析膝关节MRI诊断前交叉韧带(anterior cruciate ligament,ACL)损伤方面的应用。方法使用100名经过关节镜确认的ACL损伤患者和100名关节镜确认无ACL损伤的患者的术前MRI图像。在选取适当层面后,裁剪并用于SMRNet的训练。SMRNet对单个MRI切片上ACL损伤的概率进行最终判断。4名临床医师对相同图像进行读片诊断。结果SMRNet分类的敏感性、特异性、准确性、阳性预测值和阴性预测值分别为97.00%、94.00%、95.50%、94.17%和96.91%。医师的整体阅片情况类似,敏感性区间91.00%96.00%、特异性区间90.00%94.00%、准确性区间90.50%95.00%、阳性预测值区间90.09%94.12%、阴性预测值区间90.90%95.92%,二者差异无统计学意义(P>0.05)。结论经过训练的SMRNet模型在ACL损伤检测上超越部分临床医师,为膝关节MRI诊断提供高效可靠方法,展现深度学习在医学影像的潜力。未来,SMRNet有望成为膝关节MRI诊断的重要工具,为患者提供更精准的诊疗方案。Objective To explore the application of semi-supervised algorithm residual network(SMRNet)deep learning in computer-aided diagnosis of anterior cruciate ligament(ACL)injury in magnetic resonance images(MRI).Methods One hundred patients with arthroscopically confirmed ACL injury,and 100 patients who underwent arthroscopic examination and had no ACL injury(control group)were enrolled in the study.Preoperative MRI images of all subjects were analyzed,the appropriate layers were selected and cropped for SMRNet training.The ACL injury on a single MRI slice was determined by SMRNet and 4 senior clinicians(2 radiologists and 2 orthopedists).With the intraoperative finding as gold standard,the performance in diagnosis of ACL injury on MRI images was evaluated for two groups.Results The sensitivity,specificity,accuracy,positive predictive value and negative predictive value of SMRNet classification were 97.00%,94.00%,95.50%,94.17%and 96.91%,respectively;while the results from clinicians were similar,with sensitivity range of 91.00%-96.00%,specificity range of 90.00%-94.00%,accuracy range of 90.50%-95.00%,positive predictive value range of 90.09%-94.12%,negative predictive value range of 90.90%-95.92%,and there were no significant differences between them(P>0.05).Conclusion The diagnostic results of ACL on MRI images given by trained SMRNet model are similar to those given by senior clinicians,indicates that computer-aided diagnosis based on SMRNet deep learning is expected to become an important tool for clinical application in the future.

关 键 词：深度学习 机器学习 人工智能 前交叉韧带 磁共振成像 

分 类 号：R684[医药卫生—骨科学]