Texture Analysis and Machine Learning for Detecting Myocardial Infarction in Noncontrast Low-Dose Computed Tomography Unveiling the Invisible
Texture Analysis and Machine Learning for Detecting Myocardial Infarction in Noncontrast Low-Dose Computed Tomography Unveiling the Invisible
复制标题
DOI：
10.1097/rli.0000000000000448
复制
发表时间：
2018-06-01
影响因子：
6.7
通讯作者：
Alkadhi, Hatem
中科院分区：
文献类型：
--
作者：
Mannil, Manoj;von Spiczak, Jochen;Alkadhi, Hatem
目的：本研究的目的是测试纹理分析和机器学习是否能够在非对比增强低辐射剂量心脏计算机断层扫描（CCT）图像上检测心肌梗死（MI）。材料和方法：在这项机构审查委员会批准的回顾性研究中，我们纳入了非对比增强心电图门控低辐射剂量CCT图像数据（有效剂量，0.5 mSv），用于27例急性MI患者的钙评分（9例女性患者;平均年龄60 ± 12岁），30例慢性MI患者（8例女性患者;平均年龄，68 ± 13岁）和30例无心脏异常的受试者（9例女性患者;平均年龄，44 ± 6岁），以下称为对照组。使用徒手感兴趣区域进行左心室的纹理分析，并将纹理特征分类两次（模型I：对照与急性MI与慢性MI;模型II：对照与急性和慢性MI）。对于这两种分类，使用了6种常用的机器学习分类器：决策树C4.5（J 48），k-最近邻，局部加权学习，随机森林，顺序最小优化和采用深度学习的人工神经网络。此外，2盲，独立的读者视觉评估的非对比CCT图像的存在或不存在MI。结果：在模型I中，最好的分类结果，获得使用的k-最近邻分类器（灵敏度，69%;特异性，85%;假阳性率，0.15）。在模型II中，最好的分类结果被发现与局部加权学习分类（灵敏度，86%;特异性，81%;假阳性率，0.19）与曲线下面积从接收器操作特性分析为0.78。相比之下，这两个读者都不能识别MI的任何noncontrast，低辐射剂量CCT images.Conclusions：这项研究表明，纹理分析和机器学习的能力，在检测MI的noncontrast低辐射剂量CCT图像是不可见的放射科医生的眼睛。
Objectives: The aim of this study was to test whether texture analysis and machine learning enable the detection of myocardial infarction (MI) on non-contrast-enhanced low radiation dose cardiac computed tomography (CCT) images.Materials and Methods: In this institutional review board-approved retrospective study, we included non-contrast-enhanced electrocardiography-gated low radiation dose CCT image data (effective dose, 0.5 mSv) acquired for the purpose of calcium scoring of 27 patients with acute MI (9 female patients; mean age, 60 12 years), 30 patients with chronic MI (8 female patients; mean age, 68 13 years), and in 30 subjects (9 female patients; mean age, 44 +/- 6 years) without cardiac abnormality, hereafter termed controls. Texture analysis of the left ventricle was performed using free-hand regions of interest, and texture features were classified twice (Model I: controls versus acute MI versus chronic MI; Model II: controls versus acute and chronic MI). For both classifications, 6 commonly used machine learning classifiers were used: decision tree C4.5 (J48), k-nearest neighbors, locally weighted learning, RandomForest, sequential minimal optimization, and an artificial neural network employing deep learning. In addition, 2 blinded, independent readers visually assessed noncontrast CCT images for the presence or absence of MI.Results: In Model I, best classification results were obtained using the k-nearest neighbors classifier (sensitivity, 69%; specificity, 85%; false-positive rate, 0.15). In Model II, the best classification results were found with the locally weighted learning classification (sensitivity, 86%; specificity, 81%; false-positive rate, 0.19) with an area under the curve from receiver operating characteristics analysis of 0.78. In comparison, both readers were not able to identify MI in any of the noncontrast, low radiation dose CCT images.Conclusions: This study indicates the ability of texture analysis and machine learning in detecting MI on noncontrast low radiation dose CCT images being not visible for the radiologists' eye.