| 贾亚男,赵鹏飞,李若平,李育威,张今尧,许晓阳,朱海峰.基于冠状动脉CT血管造影深度神经网络模型检出和分类左前降支心肌桥[J].中国医学影像技术,2022,38(6):859~863 |
| 基于冠状动脉CT血管造影深度神经网络模型检出和分类左前降支心肌桥 |
| Deep neural network model based on coronary CT angiography for detection and classification of left anterior descending branch myocardial bridge |
| 投稿时间：2021-07-20 修订日期：2021-10-21 |
| DOI：10.13929/j.issn.1003-3289.2022.06.016 |
| 中文关键词: 冠状动脉疾病 心肌桥 神经网络,计算机 左前降支 |
| 英文关键词:coronary artery disease myocardial bridging neural networks, computer left anterior descending branch |
| 基金项目: |
|
| 摘要点击次数: 2247 |
| 全文下载次数: 425 |
| 中文摘要: |
| 目的 观察基于冠状动脉血管造影（CCTA）构建左前降支（LAD）心肌桥检出和分类深度神经网络模型的应用价值。方法 回顾性分析726例因胸痛不适接受冠状动脉CTA患者，由2名影像科医师诊断心肌桥并分为浅表型或纵深型。分别按7 ∶ 3及1 ∶ 2比例将心肌桥患者及非心肌桥患者归入训练集和测试集。建立基于CCTA的深度神经网络残差-长短期记忆级联网络（Resnet-LSTM）模型，用于检测及分类LAD心肌桥。采用训练集对模型进行学习训练；以影像科医师诊断结果为标准，利用测试集数据检测模型性能。结果 CCTA显示，726例中，453例存在至少1处LAD心肌桥；共检出654处心肌桥，其中561处位于LAD、93处位于左回旋支及第一对角支等。561处LAD心肌桥中，503处为浅表型、58处为纵深型。训练集含333例心肌桥患者共406处LAD心肌桥（365处浅表型和41处纵深型）和91例非心肌桥患者；测试集含120例心肌桥患者共155处LAD心肌桥（138处浅表型和17处纵深型，139处位于LAD近、中段，16处位于LAD远段）和182例非心肌桥患者。针对测试集数据，Resnet-LSTM模型检出130例存在137处LAD心肌桥（含浅表型120处、纵深型17处），误诊27例、漏诊17例，诊断敏感度为85.83%（103/120），特异度为85.16%（155/182），与医师诊断结果的一致性高（Kappa=0.70，P<0.05）；检出131处LAD近中段心肌桥，误诊2处、漏诊10处，诊断准确率为92.81%（129/139）；检出36处LAD远段心肌桥，误诊28处、漏诊8处，诊断准确率为50.00%（8/16）；将其中13处纵深型误诊为浅表型、9处浅表型误诊为纵深型，分类准确率为83.94%（115/137）。结论 基于CCTA的深度神经网络Resnet-LSTM模型用于检出和分类冠状动脉LAD心肌桥性能较好，具有一定临床应用价值。 |
| 英文摘要: |
| Objective To observe the value of the constructed deep neural network based on coronary CT angiography (CCTA) for automatic detection and classification of left anterior descending branch (LAD) myocardial bridge. Methods Data of 726 patients who underwent CCTA for chest pain and discomfort were retrospectively analyzed. The diagnosis and classification (superficial type or deep type) of myocardial bridges were performed by 2 radiologists. Patients with and without myocardial bridges were assigned to training set and test set according to the ratio of 7:3 and 1:2, respectively. Then a deep neural network residual network-long short-term memory (Resnet-LSTM) model was established based on CCTA to detect and classify LAD myocardial bridge. Data in the training set were used to train the model, while those in the test set were used to test the performances of the model according to the diagnostic results of radiologists. Results CCTA showed that, among 726 patients, 453 cases were found with at least 1 myocardial bridge in LAD, and totally 654 myocardial bridges were diagnosed, 561 in LAD and the rest 93 in the left circumflex branch or the first diagonal branch, including 503 superficial type and 58 deep type. In the training set, there were 333 cases with myocardial bridge and 91 cases without myocardial bridge, totally 406 LAD myocardial bridges including 365 superficial type and 41 deep type. In the test set, 120 cases were found with and 182 cases without myocardial bridge, and totally 155 LAD myocardial bridges were detected, including 138 superficial type and 17 deep type ones, among which 139 located at the proximal or middle part and 16 at the distal of LAD. In the test set, Resnet-LSTM model detected 137 LAD myocardial bridges in 130 cases, misdiagnosed 27 cases and missed 17 cases, with the sensitivity of 85.83% (103/120) and specificity of 85.16% (155/182), with high consistency to the radiologists' diagnostic results (Kappa=0.70, P<0.05). Resnet-LSTM model detected 131 proximal or middle LAD myocardial bridges, misdiagnosed 2 and missed 10, the diagnostic accuracy was 92.81% (129/139). Resnet-LSTM model detected 36 distal LAD myocardial bridges, misdiagnosed 28 and missed 8, and the diagnostic accuracy was 50.00% (8/16). Resnet-LSTM model correctly classified 120 LAD myocardial bridges as superficial type and 17 as deep type, misdiagnosed 13 deep type as superficial type and 9 superficial type as deep type, the classification accuracy was 83.94% (115/137). Conclusion The established deep neural network Resnet-LSTM model based on CCTA had good performances for detecting and classifying LAD myocardial bridges, hence having certain clinical application value. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
