Department of Radiology, Liyuan Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China.
College of Life Science & Technology, Huazhong University of Science and Technology, Wuhan, China.
Eur J Radiol. 2021 Mar;136:109528. doi: 10.1016/j.ejrad.2021.109528. Epub 2021 Jan 8.
The purpose of this study is to develop and evaluate a deep learning model to assist radiologists in classifying lower extremity arteries based on the degree of arterial stenosis caused by plaque in lower extremity computed tomography angiography (CTA) of patients with peripheral artery disease.
In this retrospective study, 265 patients who underwent lower-extremity CTA between January 1, 2016 and October 31, 2019 were selected. A total of 17050 axial images of iliac, femoropopliteal and infrapopliteal artery from these patients were used for the training and validation of the parallel efficient network (p-EffNet), a kind of supervised convolutional neural network, to classify the lower-extremity artery segments according to the degree of stenosis with digital subtraction angiography as reference standard. The classification results of the p-EffNet were then compared with those obtained from radiologists. Receiver operating characteristic curve (ROC) was used to evaluate the performance of the p-EffNet and accuracy, specificity, sensitivity and area under the curve (AUC) were used as measure metrics to compare the performance of the p-EffNet and that of radiologists.
The p-EffNet exhibited a good performance of 91.5 % accuracy, 0.987 AUC and 90.2 % sensitivity and 97.7 % specificity in classifying above-knee artery and 90.9 % accuracy, 0.981 AUC, 91.3 % sensitivity and 95.2 % specificity in classifying below-knee artery. When compared with human readers, for both above-knee and below-knee artery, the p-EffNet had comparable accuracy (p = 0.266 and p = 0.808, respectively) and specificity (p = 0.118 and p = 0.971, respectively) but lower sensitivity (p < 0.001 and p = 0.022, respectively).
The p-EffNet demonstrates promising diagnostic performance and has the potential to reduce the workload of radiologists and help to find the plaques that might otherwise have been missed or misjudged.
本研究旨在开发和评估一种深度学习模型，以协助放射科医生根据下肢计算机断层血管造影术（CTA）中斑块引起的动脉狭窄程度对下肢动脉进行分类，这些患者患有外周动脉疾病。
在这项回顾性研究中，选择了 2016 年 1 月 1 日至 2019 年 10 月 31 日期间接受下肢 CTA 的 265 名患者。从这些患者中总共使用了 17050 张髂动脉、股腘动脉和腘下动脉的轴向图像，以训练和验证平行高效网络（p-EffNet），这是一种监督卷积神经网络，用于根据参考标准（数字减影血管造影）对下肢动脉节段进行狭窄程度分类。然后将 p-EffNet 的分类结果与放射科医生的分类结果进行比较。接收器工作特征曲线（ROC）用于评估 p-EffNet 的性能，准确性、特异性、敏感性和曲线下面积（AUC）作为衡量指标，用于比较 p-EffNet 和放射科医生的性能。
p-EffNet 在分类膝上动脉时表现出良好的性能，准确率为 91.5%，AUC 为 0.987，敏感性为 90.2%，特异性为 97.7%，在分类膝下动脉时准确率为 90.9%，AUC 为 0.981，敏感性为 91.3%，特异性为 95.2%。与人类读者相比，p-EffNet 在膝上动脉和膝下动脉的分类中均具有相当的准确性（p=0.266 和 p=0.808）和特异性（p=0.118 和 p=0.971），但敏感性较低（p<0.001 和 p=0.022）。
p-EffNet 显示出有希望的诊断性能，有可能减轻放射科医生的工作量，并有助于发现可能被遗漏或误诊的斑块。
