基于双流神经网络的颈动脉粥样硬化斑块稳定性区分方法 Method of Distinguishing Stability of Carotid Plaque Based on Two-stream Neural Network

目的 训练双流神经网络自动区分颈动脉粥样硬化斑块的稳定性。 方法 使用颈动脉内膜剥脱术后经病理证实的115例稳定颈动脉粥样硬化斑块患者和110例易损颈动脉粥样硬化斑块患者的844段超声造影视频，将其中744段视频作为训练集，训练双流神经网络，得到在训练集上区分效果最佳的神经网络。将剩余的100段视频作为测试集，测试该神经网络自动区分颈动脉粥样硬化斑块稳定性的价值。 结果 神经网络在训练集上区分颈动脉斑块稳定性的准确率、敏感度、特异度、阳性预测值、阴性预测值、阳性似然比、阴性似然比分别为93%、87%、97%、96%、90%、29和0.13，在测试集上相应的结果分别为80%、70%、90%、88%、75%、7和0.33。受试者工作特征曲线上，训练集和测试集中双流神经网络判断斑块易损性的曲线下面积分别为0.99和0.84，均P＜0.001。 结论 利用已知病理结果的超声造影视频，将其输入到双流神经网络进行训练，能得到较好的自动区分颈动脉粥样硬化斑块稳定性的模型。 Objective To train two-stream neural network to distinguish the stability of carotid plaques. Methods 844 contrast-enhanced ultrasound videos were used in the experiment. They were from 115 patients with stable carotid plaques and 110 patients with vulnerable carotid plaques verified by pathology after CEA. 744 videos were used as training set to train two-stream network, to find the neural network segment having optimal recognition effect. The left 100 videos were used as test set to distinguish the stability of carotid plaque. Results Accuracy, sensitivity, specificity, positive predictive value, negative predictive value, positive likelihood ratio and negative likelihood ratio in training set were 93%, 87%, 97%, 96%, 90%, 29 and 0.13, respectively. The corresponding results in test set were 80%, 70%, 90%, 88%, 75%, 7 and 0.33, respectively. Area under the receiver operating characteristic curve for training set and test set were 0.99 and 0.84 (both P