基于深度学习训练上腹部DCE-MRI扫描期相分类模型的可行性  

The feasibility of training classification model of different phases of upper abdominal DCE-MRI images based on deep learning method

作　　者：额·图娅 郭小超[1] 王可[1] 黄嘉豪 王祥鹏 张晓东[1] 王霄英[1] E·Tu-ya;GUO Xiao-chao;WANG Ke(Department of Radiology,the First Affiliated Hospital of Beijing University,Beijing 100034,China)

机构地区：[1]北京大学第一医院医学影像科,北京100034 [2]北京赛迈特锐医疗科技有限公司,北京100034

出　　处：《放射学实践》2021年第8期1052-1058,共7页Radiologic Practice

摘　　要：目的:探讨基于深度学习算法建立的上腹部DCE-MRI图像自动分类模型的立场应用价值。方法:回顾性搜集417例患者上腹部DCE-MRI扫描不同期相的共1330组图像数据。由两位专家将所有图像按平扫、动脉早期、动脉晚期和门静脉-延迟期进行分类。将1330组的数据随机分为训练集(train set,n=1118),调优集(validation set,n=108)和测试集(test set,n=104)。训练3D-ResNet模型对图像的扫描期相进行分类,应用混淆矩阵(confusion matrix)评价模型的分类预测效能。结果:在训练集、调优集及测试集中总体预测符合率分别为99.9%(1117/1118)、99.1%(107/108)和99.0%(103/104)。训练集及调优集中动脉晚期的预测符合率分别为99.5%(193/194)和90.9%(10/11),平扫、动脉早期及门静脉-延迟期的预测符合率均为100%。测试集中各期相图像分类符合率:平扫97.5%(39/40)、动脉早期100%(14/14)、动脉晚期100%(8/8)、门静脉-延迟期100%(42/42)。结论:基于深度学习方法训练的分类模型对DCE-MRI各期相图像的分类预测效能良好,有利于工作流程的优化及后续对接AI诊断模型。Objective:To develop a model based on deep learning to classify DCE-MRI images of upper abdomen in multi-phase dynamic enhanced scan.Methods:417 cases undergone upper abdominal pre-contrast and DCE-MRI examinations were retrospectively included in this study,and 1330 groups was established by putting the single phase images in each examination into one group.The images were labeled as non-contrast,early arterial phase,late arterial phase,and portal vein-delayed phase by two experienced radiologists.The 1330 series DCE-MRI images were randomly divided as train set(n=1118),validation set(n=108)and test set(n=104).A classification model was trained using 3D-ResNet network,and the classification efficacy of the model was evaluated by using confusion matrix.Results:The overall accuracy of the model in the classification of upper abdominal DCE-MRI images was as following:99.9%(1117/1118)in the train set,99.1%(107/108)in the validation set,and 99.0%(103/104)in the test set.In the train set and validation set,the prediction accuracy of the late arterial phase was 99.5%(193/194)and 90.0%(10/11),and 100%for all the other phases.In the test set,the prediction accuracy was as following:non-contrast phase 97.5%(39/40),early arterial phase 100%(14/14),late arterial phase 100%(8/8),and portal vein-delayed phase 100%(42/42).Conclusion:The classification model for abdominal DCE-MRI images has good performance,which is beneficial to the quality control and the subsequent implementation of AI diagnosis models.

关 键 词：磁共振成像 对比增强扫描 深度学习 图像分类 质量控制 

分 类 号：R445.2[医药卫生—影像医学与核医学]