摘要
目的:探索利用深度学习方法在建立颈椎病MR诊断模型的可行性.方法:回顾性搜集本院2020年10月-2023年3月诊断为颈椎病患者的MR图像514例,使用已有颈椎分割模型在轴面T2WI上分别预测硬膜囊、脊髓、椎间盘、后纵韧带和黄韧带,在矢状面T1WI和T2WI上预测颈椎椎体和椎间盘.由一位低年资放射科医生(阅片经验2年)修改标注,另一位高年资放射科医生(阅片经验≥15年)对低年资医师的标注进行复核.按照颈椎病的不同诊断要点分别进行3D或2D深度学习分类模型训练,包括①颈椎椎体增生模型;②颈椎椎体滑脱模型;③颈椎间盘突出分类模型;④后纵韧带增厚模型;⑤黄韧带增厚模型.将模型输出结果导入R软件进行混淆矩阵分析及ROC曲线绘制,采用正确率、灵敏度、特异度、阳性预测值、阴性预测值以及ROC曲线下面积等评价5种模型的分类效能.结果:5种分类模型中诊断效能最好的是颈椎间盘突出分类模型,正确率0.90,灵敏度0.95,特异度0.85,ROC曲线下面积0.982.颈椎椎体增生和滑脱的正确率分别为0.81和0.80,灵敏度为0.74和0.76,特异度为0.84和1.00,ROC曲线下面积分别为0.855和0.905.后纵韧带和黄韧带增厚的模型正确率分别为0.82和0.77,灵敏度为0.78和0.84,特异度为0.86和0.70,ROC曲线下面积分别为0.902和0.929.结论:本部分研究采用深度学习方法建立了颈椎病MR的自动分类诊断模型,对颈椎椎体增生、滑脱、椎间盘突出、后纵韧带及黄韧带增厚进行了分类模型训练,证明深度学习方法可以用于颈椎病MR的辅助诊断,为未来进一步探索建立颈椎病MR自动诊断模型及结构化报告的植入奠定了基础.
Abstract
Objective:To explore the feasibility of using deep learning methods in establishing MRI diagnostic models for cervical spondylosis.Methods:A retrospective collection of 514 MR images of patients were diagnosed with cervical spondylosis in our hospital from October 2020 to March 2023.Use existing cervical spine segmentation models to predict the subarachnoid space,spinal cord,inter-vertebral disc,posterior longitudinal ligament,and ligamentum flavum on axial T2WI,and predict the cervical vertebral bodies and intervertebral disc on sagittal T1 WI and T2 WI.A junior radiologist(with 2 years of experience)revised the labeling,and another senior radiologist(with≥15 years of experi-ence)reviewed the labels.According to the different radiologic sign of cervical spondylosis,3D or 2D U-net deep learning classification model training is carried out,including cervical vertebral hyperplasia model,cervical spondylolisthesis model,cervical disc herniation classification model,thickness of pos-terior longitudinal ligament model,and thickness of ligament flavum model.The output results of the models were imported into R programming software for confusion matrix analysis and ROC curve drawing.And the classification performance of the five models(accuracy,sensitivity,specificity,posi-tive predictive value,negative predictive value,and area under the ROC curve)was evaluated.Results:Among the five classification models,the cervical disc herniation classification model had the best di-agnostic performance,with a correct rate of 0.90,sensitivity of 0.95,specificity of 0.85,and area under the ROC curve of 0.982.The correct rate of cervical vertebral hyperplasia and spondylolisthesis also reached 0.81 and 0.80,the sensitivity was 0.74 and 0.76,the specificity was 0.84 and 1.00,and the area under the ROC curve was 0.855 and 0.905 respectively.The correct rates of the models for thickening of the posterior longitudinal ligament and ligamentum flavum were 0.82 and 0.77,respectively,with sensitivities of 0.78 and 0.84,specificities of 0.86 and 0.70,and areas under the ROC curve of 0.902 and 0.929,respectively.Conclusion:In the study,deep learning method based on MRI was used to establish the automatic classification diagnosis model of cervical spondylosis,and the classification model was trained for cervical vertebral body hyperplasia,slippage,intervertebral disc herniation,thickness of posterior longitudinal ligament and ligamentum flavum.The study proved that deep learning method based on MRI can be used in the computer aided diagnosis of cervical spondylosis,which laid the groundwork and foundation of the development of structured reports of cervical spine.
NETL
NSTL
万方数据