基于多模态超声图像的深度学习在鉴别诊断肝内胆管细胞癌与肝细胞癌中的应用

Application of deep learning based on multimodal ultrasound images in differential diagnosis of ICC and HCC

目的探讨基于单独二维超声图像(B-mode)、超声造影(CEUS)图像、联合B-mode及CEUS图像所构建的不同卷积神经网络(CNN)模型对鉴别诊断肝内胆管细胞癌(ICC)与肝细胞癌(HCC)的能力。方法回顾性分析2011年1月—2018年5月在四川大学华西医院超声科进行肝脏占位性病变超声检查的病例,对所纳入的病例超声图像进行肿瘤范围感兴趣区标记、数据增强后,根据3种不同组合方式的图像输入进行CNN模型训练,从而建立鉴别诊断ICC与HCC的不同预测模型,并对模型的诊断效能进行评价。结果最终共有305例ICC和945例HCC被纳入研究,共计超声图像7427幅。结果显示,基于单独B-mode图像所建立的CNN模型诊断效能均较低,其中VGG16的诊断效能相对较高,优于其它两个模型(P<0.05),ROC曲线下面积(AUC)为0.672。基于CEUS图像的多图VGG和多图Inception预测模型对鉴别ICC与HCC的AUC分别为0.941和0.934(P=0.803),准确性分别为87.0%和89.0%,与基于联合B-mode及CEUS图像的两种CNN模型诊断效能相近(P>0.05)。结论基于多模态超声图像所建立的CNN模型能较好的用于鉴别诊断ICC与HCC。

Objective The ability of different convolutional neural networks(CNN)models based on single conventional 2D ultrasound images,CEUS images,combined conventional 2D ultrasound and CEUS images to distinguish intrahepatic cholangiocarcinoma(ICC)from hepatocellular carcinoma(HCC)was preliminarily discussed.Methods According to the inclusion and exclusion criteria,the cases of focal liver lesions undergoing ultrasound examination in the Department of Ultrasound,West China Hospital,Sichuan University from January 2011 to May 2018 were retrospectively analyzed.After region of interest labeling and data augmentation,the CNN model was trained according to the three different combinations of input of image data,so as to establish different prediction models for differential diagnosis of ICC and HCC,and evaluate the diagnostic performance of the models.Results A total of 305 patients with ICC and 945 patients with HCC were finally included in the study,with a total of 7427 ultrasound images.The results showed that the diagnostic performance of the CNN models based on B-mode images alone were all low,and the diagnostic performance of VGG16 was relatively high,but the AUC was only 0.672.The AUC of the multi-image VGG and multi-image Inception prediction models based on CEUS images for discriminating ICC from HCC were 0.941 and 0.934(P=0.803),respectively,with accuracies of 87.0%and 89.0%,respectively.It is similar to the CNN models based on the combined B-mode and CEUS images.Conclusion The CNN models based on multimodal ultrasound images can be better used for the differential diagnosis of ICC and HCC.