Hybrid 3D/2D Convolutional Neural Network for Hemorrhage Evaluation on Head CT.
Hybrid 3D/2D Convolutional Neural Network for Hemorrhage Evaluation on Head CT.
复制标题
DOI：
10.3174/ajnr.a5742
复制
发表时间：
2018-09
期刊：
影响因子：
--
通讯作者：
Chow D
中科院分区：
文献类型：
--
作者：
Chang PD;Kuoy E;Grinband J;Weinberg BD;Thompson M;Homo R;Chen J;Abcede H;Shafie M;Sugrue L;Filippi CG;Su MY;Yu W;Hess C;Chow D
卷积神经网络(CNN)是一种强大的图像识别技术。本研究评价了一种最适合在非增强CT(NCCT)上检测和量化实质内(IPH)、硬膜外/硬膜下(EDH/SDH)和蛛网膜下腔(SAH)出血的CNN。这项研究分两个阶段进行。首先，使用一家机构在2017年1月1日至2017年7月31日期间获得的所有NCC的培训队列来开发和交叉验证用于出血评估的定制混合3D/2D MASK R-CNN架构。其次，将训练好的网络前瞻性地应用于2018年2月1日至2018年2月28日期间在自动推理管道中从急诊科订购的所有NCCT。根据完整和平衡的数据集评估出血检测的准确性、AUC、敏感性、特异性、PPV和NPV，并进一步根据出血类型和大小进行分层。量化采用Dice评分系数和Pearson相关系数。这项研究共使用了10,159次考试训练队列(512,598张图像，901/8.1%的出血)和862次考试测试队列(23,668张图像，82/12%的出血)。在训练队列交叉验证中，出血检测的准确性、曲线下面积、敏感性、特异性、PPV值和NPV值分别为0.975、0.983、0.971、0.975、0.793和0.997，而前瞻性测试集的准确性、曲线下面积、敏感度、特异性、PPV值和NPV值分别为0.970、0.981、0.951、0.973、0.829和0.993。IPH、EDH/SDH和SAH的DICE得分分别为0.931、0.863和0.772。定制的深度学习工具可以准确地检测和量化NCCT上的出血。从急诊科订购的未来NCCT的高性能表明，拟议的深度学习工具在临床上是可行的。
Convolutional neural networks (CNN) are a powerful technology for image recognition. This study evaluates a CNN optimized for the detection and quantification of intraparenchymal (IPH), epidural/subdural (EDH/SDH) and subarachnoid (SAH) hemorrhages on non-contrast CT (NCCT). This study was performed in two phases. First, a training cohort of all NCCTs acquired at a single institution between January 1, 2017 and July 31, 2017 was used to develop and cross-validate a custom hybrid 3D/2D mask R-CNN architecture for hemorrhage evaluation. Second, the trained network was applied prospectively to all NCCTs ordered from the emergency department between February 1, 2018 and February 28, 2018 in an automated inference pipeline. Hemorrhage detection accuracy, AUC, sensitivity, specificity, PPV, and NPV was assessed for full and balanced datasets, and further stratified by hemorrhage type and size. Quantification was assessed by Dice score coefficient and Pearson correlation. A total of 10,159-exam training cohort (512,598 images; 901/8.1% hemorrhages) and 862-exam test cohort (23,668 images; 82/12% hemorrhags) were used in this study. Accuracy, area under the curve, sensitivity, specificity, PPV, and NPV for hemorrhage detection were 0.975, 0.983, 0.971, 0.975, 0.793, and 0.997 upon training cohort cross-validation, and 0.970, 0.981, 0.951, 0.973, 0.829, and 0.993 for the prospective test set. Dice scores for IPH, EDH/SDH, and SAH were 0.931, 0.863 and 0.772, respectively. A customized deep learning tool is accurate in detection and quantification of hemorrhage on NCCT. Demonstrated high performance on prospective NCCTs ordered from the emergency department suggests the clinical viability of the proposed deep learning tool.