摘要
目的建立并评价基于机会性CT检查的骨质疏松筛查分类和骨密度值预测的深度学习神经网络模型。方法以定量计算机断层扫描(Quantitative Computed Tomography,QCT)骨密度测定为标准,将199例机会性CT检查数据用于建立密集卷积网络的深度学习神经网络的骨密度二分类模型和骨密度值预测回归模型,以五折交叉验证和随机分组的方法进行测试,并以来自不同设备的42例机会性CT检查病例进行独立测试,计算和评价模型的性能参数。结果受试者工作特征(Receiver Operating Characteristic,ROC)曲线显示:骨密度二分类模型的测试集和独立测试集的ROC曲线下面积均值分别为0.974、0.938,测试集的F1得分、召回率、精准度、特异性、准确度均≥0.91,独立测试集的上述评价参数均>0.862。在训练集、测试集和独立测试集上,骨密度值预测回归模型的平均绝对误差分别为1.42、8.52和13.89,均方根误差分别为1.93、10.80、20.36,预测值与QCT骨密度值呈极强正相关。结论基于机会性CT检查的深度学习神经网络模型对骨密度正常和降低具有较强的分类能力,且可较准确地预测骨密度值,避免多余的辐射风险,减少时间、经济消耗,有效扩大骨质疏松筛查的范围。
Objective To establish and evaluate a deep learning neural network model for osteoporosis screening classification and bone density prediction based on opportunistic CT.Methods The quantitative computed tomography(QCT)bone density measurement was used as the standard,199 cases of opportunistic CT data were selected to establish deep learning neural networks for densely convolutional networks models for bone density binary classification and bone density value regression.Five-fold cross�validation and random grouping methods were used for testing,and the performance parameters of the models were calculated and evaluated using an independent test set of 42 opportunistic CT cases from different devices.Results The receiver operating characteristic(ROC)curves showed that the mean area under curve for the bone density binary classification model in the testing set and the independent test set were 0.974 and 0.938,respectively.The F1 score,recall,precision,specificity,and accuracy of the testing set were all greater than or equal to 0.91,while the aforementioned evaluation parameters for the independent test set were all greater than 0.862.The mean absolute error of the bone density value prediction regression model in the training set,testing set,and independent test set were 1.42,8.52,and 13.89,respectively,and the root mean square error were 1.93,10.80,and 20.36,respectively.The predicted values showed a strong positive correlation with QCT bone density values.Conclusion The deep learning neural network model based on opportunistic CT represent strong classification ability for normal and decreased bone density and can accurately predict bone density values,avoid unnecessary radiation risks and reduce time and economic consumption,which is conducive to effectively expanding the scope of osteoporosis screening.
作者
彭涛
曾小辉
李洋
李曼
蒲冰洁
植彪
王永芹
PENG Tao;ZENG Xiaohui;LI Yang;LI Man;PU Bingjie;ZHI Biao;WANG Yongqin(Department of Radiology,Affiliated Hospital of Chengdu University,Chengdu Sichuan 610081,China;Department of Research,Shanghai United Imaging Intelligence Co.,Ltd.,Shanghai 200000,China)
出处
《中国医疗设备》
2024年第2期57-62,74,共7页
China Medical Devices
基金
成都市卫生健康委员会医学科研课题(2021036,2021045)。
关键词
骨质疏松筛查
机会性CT
人工智能
骨密度
卷积神经网络
osteoporosis screening
opportunistic CT
artificial intelligence
bone density
convolutional neural network
