基于双路卷积神经网络在甲状腺结节良恶性鉴别诊断中的初步研究

A preliminary study on differential diagnosis of benign and malignant thyroid nodules based on dual-channel convolutional neural network

目的:考虑到甲状腺结节的三维立体形状,超声检查一般通过横切面和纵切面综合观察甲状腺结节的特征,为此我们提出了双路卷积神经网络的甲状腺结节识别模型,旨在提高甲状腺良恶性结节的鉴别诊断水平。方法:从内蒙古医科大学附属医院超声数据库中选取经手术或细针穿刺细胞学检查(FNAC)病理证实的甲状腺结节1105枚,每个结节均提供横切图与纵切图。选取884枚结节(其中恶性结节680枚,良性结节204枚)用于深度学习模型的训练,余221枚结节(良性结节59枚,恶性结节162枚)用于测试。模型考虑甲状腺结节横切图与纵切图的特征,构建双路卷积神经网络结构。首先分别训练两路子网络,其中一条以结节横切图像输入(CNN1),另一条以相应结节的纵切图像输入(CNN2),分别用于测试,然后通过特征融合层将两类特征相加为融合特征,并利用全连接层对甲状腺结节的良恶性进行识别。以术后病理学结果为金标准,分析双路卷积神经网络与两条单路卷积神经网络模型的诊断效能及与病理结果之间的一致性。结果:双路卷积神经网络模型诊断甲状腺结节的灵敏度、特异度、准确度分别为95.68%、84.75%、92.76%,均优于单路卷积神经网络模型(均P<0.05);CNN1与CNN2相比,灵敏度、特异度及准确度均无显著性差异(P>0.05)。双路卷积神经网络模型与病理诊断的一致性好(Kappa值=0.813,P<0.05);CNN1、CNN2与病理诊断的一致性一般(Kappa值=0.460、Kappa值=0.521,P<0.05)。结论:双路卷积神经网络模型理论上能够更全面的提取甲状腺结节的图像特征,更拟合超声检查中的多切面扫查,此方法应用于甲状腺结节良恶性鉴别诊断中是可行的。

Objective:In consideration of the three-dimensional shape of thyroid nodule,ultrasonic examination is generally used to comprehensively observe the characteristics of thyroid nodule through transverse and longitudinal sections.Therefore,we proposed a dual-channel convolutional neural network model for thyroid nodule recognition to improve the level of differential diagnosis of benign and malignant thyroid nodule.Methods:A total of 1105 thyroid nodules confirmed by surgery or fine-needle aspiration cytology(FNAC)pathology were selected from the hospital ultrasound database.The transverse and longitudinal sections were provided for each nodule.A total of 884 nodules(including 680 malignant nodules and 204 benign nodules)were selected for deep learning model training,and 221 nodules(59 benign nodules and 162 malignant nodules)were used for testing.The model took into account the features of transverse and longitudinal images of thyroid nodule,and we constructed a dual convolutional neural network structure.First,two subnetworks were trained respectively,one of which was input with crosscutting image(CNN1),the other was input with longitudinal cutting image of corresponding nodules(CNN2),respectively for testing.Then,the two kinds of features were added into fusion features through feature fusion layer,and the full connection layer was used to identify benign and malignant thyroid nodules.Using postoperative pathological results as the gold standard,the consistency between the diagnostic efficacy and pathological results of dual-channel convolutional neural network and single-channel convolutional neural network was analyzed.Results:The sensitivity,specificity and accuracy of dual-channel convolutional neural network model were 95.68%,84.75%and 92.76%,respectively,which were better than single-channel convolutional neural network model(all P<0.05);There were no significant differences in sensitivity,specificity and accuracy between CNN1 and CNN2(P>0.05).The dual-channel convolutional neural network model was consistent with patho-logical diagnosis(Kappa=0.813,P<0.05).CNN1 and CNN2 were generally consistent with pathological diagnosis(Kappa=0.460,Kappa=0.521,P<0.05).Conclusion:The dual-channel convolutional neural network model proposed can theoretically extract the image features of thyroid nodule more comprehensively,and better fit the multi-section scan in clinical ultrasound work.It is feasible to apply this method to the clinical ultrasonic differential diagnosis of thyroid nodule.