基于多中心心脏磁共振数据的深度学习自动评估心肌梗死后矛盾性搏动的方法
本前瞻性研究纳入401名参与者(311例患者和90名年龄匹配志愿者),建立二维UNet左心室分割模型和DCNN分类模型,利用2腔和3腔电影图像识别孤立前壁梗死直接PCI再灌注后左心室矛盾性搏动。DCNN模型在训练、内部和外部测试队列的AUC分别为0.97、0.91和0.83(p < 0.001),鉴别性能优于培训中医生(p < 0.05),2.5维多视图模型效率最高且诊断敏感度最高。
要点
- DCNN模型基于2腔和3腔电影图像自动识别左心室矛盾性搏动,外部测试AUC为0.83。
- 2.5维多视图模型结合收缩末期和舒张末期图像,效率优于3D模型且诊断敏感度最高。
- 模型鉴别性能优于培训中医生(p < 0.05),可客观评估再灌注后矛盾性搏动。
核心结论
- 整合2腔和3腔CMR图像的2.5维多视图DCNN模型能高效识别孤立前壁AMI直接PCI再灌注后左心室矛盾性搏动,且诊断敏感度最高。
研究设计与样本
- 前瞻性研究,纳入401名参与者(311例患者和90名年龄匹配志愿者),建立二维UNet左心室分割模型和DCNN分类模型,使用2D和3D ResNet提取2腔和3腔图像特征,以Dice评分评估分割模型,以ROC曲线和混淆矩阵评估分类模型,用DeLong法比较培训中医生与DCNN模型的AUC。
诊断标准与关键数值
- DCNN模型检测矛盾性搏动的AUC在训练队列为0.97、内部测试队列为0.91、外部测试队列为0.83(p < 0.001);2.5维模型(收缩末期和舒张末期图像联合2腔和3腔图像)效率高于3D模型;DCNN模型鉴别性能优于培训中医生(p < 0.05)。
临床应用与阅片要点
- 作者结论认为,与单独使用2腔或3腔图像或3D多视图训练的模型相比,2.5维多视图模型能更高效地结合2腔和3腔信息并获得最高诊断敏感度。
局限性
- 原文未报告局限性
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
英文原文(摘要)· 点击展开
OBJECTIVE: The current study aimed to explore a deep convolutional neural network (DCNN) model that integrates multidimensional CMR data to accurately identify LV paradoxical pulsation after reperfusion by primary percutaneous coronary intervention with isolated anterior infarction.
METHODS: A total of 401 participants (311 patients and 90 age-matched volunteers) were recruited for this prospective study. The two-dimensional UNet segmentation model of the LV and classification model for identifying paradoxical pulsation were established using the DCNN model. Features of 2- and 3-chamber images were extracted with 2-dimensional (2D) and 3D ResNets with masks generated by a segmentation model. Next, the accuracy of the segmentation model was evaluated using the Dice score and classification model by receiver operating characteristic (ROC) curve and confusion matrix. The areas under the ROC curve (AUCs) of the physicians in training and DCNN models were compared using the DeLong method.
RESULTS: The DCNN model showed that the AUCs for the detection of paradoxical pulsation were 0.97, 0.91, and 0.83 in the training, internal, and external testing cohorts, respectively (p < 0.001). The 2.5-dimensional model established using the end-systolic and end-diastolic images combined with 2-chamber and 3-chamber images was more efficient than the 3D model. The discrimination performance of the DCNN model was better than that of physicians in training (p < 0.05).
CONCLUSIONS: Compared to the model trained by 2-chamber or 3-chamber images alone or 3D multiview, our 2.5D multiview model can combine the information of 2-chamber and 3-chamber more efficiently and obtain the highest diagnostic sensitivity.
CLINICAL RELEVANCE STATEMENT: A deep convolutional neural network model that integrates 2-chamber and 3-chamber CMR images can identify LV paradoxical pulsation which correlates with LV thrombosis, heart failure, ventricular tachycardia after reperfusion by primary percutaneous coronary intervention with isolated anterior infarction.
KEY POINTS: • The epicardial segmentation model was established using the 2D UNet based on end-diastole 2- and 3-chamber cine images. • The DCNN model proposed in this study had better performance for discriminating LV paradoxical pulsation accurately and objectively using CMR cine images after anterior AMI compared to the diagnosis of physicians in training. • The 2.5-dimensional multiview model combined the information of 2- and 3-chamber efficiently and obtained the highest diagnostic sensitivity.
引用
- Chen, Wu, An, et al. A deep learning method for the automated assessment of paradoxical pulsation after myocardial infarction using multicenter cardiac MRI data. Eur Radiol. 2023. PMID: 37389610. doi:10.1007/s00330-023-09807-6
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。