利用深度学习实现腹盆部CT图像范围及期相分类:临床验证研究

Deep learning for classification of range and phase of abdominal and pelvic CT scanning:a prospectivevalidation study in clinical workflow

目的:探讨基于深度学习的分类模型对腹盆部CT图像范围及期相进行自动分类的可行性。方法:回顾性搜集本院2019年10月14日-2019年10月18日PACS中连续416例患者的腹盆部CT图像(数据集A)。按照扫描范围分为腹部、腹盆部、盆部三类,按照扫描期相分为平扫、动脉期、门静脉期、延迟期和排泄期五类。以3D-ResNet为基础架构,训练CT图像范围及期相的分类模型。利用该模型预测2020年1月1日-2020年1月3日本院连续657例患者的腹盆部CT图像(数据集B)。以影像医师的分类结果为金标准,采用混淆矩阵评价模型的分类效能。结果:在数据集B中,扫描范围分类模型在腹部、腹盆部和盆部的符合率分别为95.7%(243/254)、98.4%(362/368)和94.3%(33/35)。对数据集B中的腹部图像进行分析,扫描期相分类模型在平扫、动脉期和门静脉期的符合率分别为100.0%(77/77)、97.6%(82/84)和100.0%(11/11);对数据集B中腹盆部图像进行分析,扫描期相分类模型在平扫、动脉期、门静脉期、延迟期和排泄期的符合率分别为96.6%(144/149)、100.0%(9/9)、100.0%(106/106)、66.7%(44/66)和100.0%(32/32);对数据集B中盆部图像分析,扫描期相分类模型在平扫、门静脉期、延迟期和排泄期的符合率分别为100.0%(13/13)、70.0%(7/10)、88.9%(8/9)和100.0%(1/1)。结论:通过深度学习模型建立腹盆部CT图像性质分类模型的准确性基本可达到临床要求。

Objective:To explore the feasibility of a classification model established by deep learning for classifying the scan range and dynamic contrast enhanced phase of abdominal and pelvic CT images.Methods:The abdominal and pelvic CT images of 416 consecutive patients were retrospectively collected from Oct 14,2019 to Oct 18,2019(dataset A).According to the scanning range,all subjects were divided into three groups:abdomen,abdomen and pelvis,pelvis.And according to the scanning phase,they were divided into five groups:plain scan,arterial phase,portal venous phase,delayed phase and exctretory phase.The CT image range and phase classification model were trained based on 3D-ResNet.The model was used to predict 657 consecutive CT images of abdomen and pelvis from January 1,2020 to January 3,2020(Dataset B).The classification results of radiologist were taken as the gold standard,the confusion matrix was used to evaluate the classification efficiency of the model.Results:In Dataset B,the accuracy of the scanning range classification model in the abdomen,abdomen and pelvis,pelvis was 95.7%(243/254),98.4%(362/368)and 94.3%(33/35),respectively.In the abdomen images of Dataset B,the accuracy of plain scan,arterial and portal venous phase was 100.0%(77/77),97.6%(82/84)and 100.0%(11/11),respectively.In the abdomen and pelvis images of Dataset B,the accuracy of plain scan,arterial phase,portal venous phase,delayed phase and exctretory phase was 96.6%(144/149),100.0%(9/9),100.0%(106/106),66.7%(44/66)and 100.0%(32/32),respectively.In the pelvis images of Dataset B,the accuracy of pelvic scan phase classification model in plain scan,portal venous phase,delayed phase and exctretory phase was 100.0%(13/13),70.0%(7/10),88.9%(8/9)and 100.0%(1/1),respectively.Conclusion:The accuracy of abdominal and pelvic CT image classification model established by deep learning is accurate and can meet the cli-nical requirements.