基于卷积神经网络的超声影像肝癌自动分类  

Automatic Classification of Liver Cancer in B-ultrasound and Contrast-enhanced Ultrasound Based on Convolution Neural Network

作　　者：李睿[1,2,3] 许祥丛 林静怡 黄良汇 曾亚光[1,2] 郑玮[4] 陈广义[3] 王雪花 韩定安[1,2] LI Rui;XU Xiang-Cong;LIN Jing-Yi;HUANG Liang-Hui;ZENG Ya-Guang;ZHENG Wei;Chen Guang-Yi;WANG Xue-Hua;HAN Ding-An(School of Physics and Optoelectronic Engineering,Foshan University,Foshan 528225,China;Guangdong-Hong Kong-Macao Joint Laboratory for Intelligent Micro-Nano Optoelectronic,Foshan University,Foshan 528225,China;School of Mechatronic Engineering and Automation,Foshan University,Foshan 528225,China;Sun Yat-sen University Cancer Center,Guangzhou 510062,China)

机构地区：[1]佛山科学技术学院物理与光电工程学院,佛山528225 [2]佛山科学技术学院,粤港澳智能微纳光电技术联合实验室,佛山528225 [3]佛山科学技术学院机电工程与自动化学院,佛山528225 [4]中山大学肿瘤防治中心,广州510062

出　　处：《生物化学与生物物理进展》2023年第3期668-675,共8页Progress In Biochemistry and Biophysics

基　　金：广东省重点领域研究与发展计划(2020B1111040001);国家自然科学基金(61805038,62075042,61705036,61771139);粤港澳智能微纳光电技术联合实验室研究基金(2020B1212030010)资助项目。

摘　　要：目的针对从原发性肝癌中检测肝细胞癌(HCC)的灵敏度不高和诊断结果高度依赖放射科医生的专业性和临床经验,本文利用深度卷积神经网络(CNN)的方法自动学习B超和超声造影(CEUS)图像中的特征信息,并实现对肝癌的分类。方法建立并验证基于CNN的多个二维(2D)和三维(3D)分类模型,分别对116例患者(其中100例HCC和16例非HCC)的B超和CEUS影像进行定量分析,并对比分析各个模型的分类性能。结果实验结果表明,3D-CNN模型的各方面性能指标都优于2D-CNN模型,验证了3D-CNN模型能同时提取肿瘤区域的2D影像特征及血流时间动态变化特征,比2D-CNN模型更适用于HCC与非HCC分类。其中3D-CNN模型的AUC、准确率和敏感度值最高,分别达到了85%、85%和80%。此外,由于HCC和非HCC样本不均衡,通过扩充非HCC样本的数量可以提升网络的分类性能。结论本文提出的3D-CNN模型能够实现快速、准确的肝癌分类,有望应用于辅助临床医师诊断与治疗肝癌。Objective Primary liver cancer is a common malignant tumor,seriously threatening people’s life and health.According to the differences in pathogenesis,treatment and prognosis,primary liver cancer can be divided into hepatocellular carcinoma(HCC),intrahepatic cholangiocarcinoma(ICC)and other rare types.Among which HCC accounts for 85%-90%.HCC is usually treated by transcatheter chemoembolization(TACE)or minimally invasive ablation,with good prognosis.While ICC and HCC-ICC mixed type have a high degree of malignancy and are generally treated by surgical resection or liver transplantation,with poor prognosis.In order to improve the diagnostic accuracy of HCC patients,primary liver cancer is usually clinically divided into HCC and non-HCC categories,that is non-HCC includes ICC,HCC-ICC mixed type and other rare tumors.Therefore,accurate screening of HCC from liver cancer lesions is of great clinical significance for the treatment of HCC patients.However,due to the high heterogeneity of tumors,the shape,texture,location and blood flow of liver lesions show complexity and diversity in B-ultrasound and contrast-enhanced ultrasound(CEUS)images.Radiologists need to rely on the naked eyes to obtain multidimensional information at the same time,and evaluate diseases according to different characteristics,which requires high level of expertise and clinical experience.Diagnosis results depend on personal subjective factors,which may lead to some HCC mixed into non-HCC categories,and the detection sensitivity of HCC is not high.In this paper,deep convolutional neural network is used to automatically learn the characteristic information of B-ultrasound and CEUS images,and realize the classification of liver cancer.Methods Multiple 2D(VGG,ResNet,DenseNet)and 3D(3D-CNN,Res3D,Dense3D)classification models based on convolutional neural network(CNN)were established and validated,and the B-ultrasound and CEUS images of 116 patients(including 100 HCC and 16 non-HCC patients)were quantitatively analyzed,and the classification perform

关 键 词：深度学习 卷积神经网络 肝癌分类 超声影像 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术] R735.7[医药卫生—肿瘤]