基于deep CNN的Faster R-CNN在肝脏肿瘤超声图像中的初步应用  

Preliminary Application of Faster R-CNN Based on Deep CNN in Ultrasound Images of Liver Masses

作　　者：管海涛[1] 赵金丽[2] Xi Pengcheng Guan Haitao;Zhao Jinli;Xi Pengcheng(Department of Ultrasound,Nantong Third People′s Hospital,Nantong,Jiangsu 226006;China 2.Imaging Department,Affiliated Hospital of Nantong University,Nantong,Jiangsu 226001;China 3.Digital Technologies Research Center,National Research Council of Canada,Ottawa,Ontario K1A0R6,Canada)

机构地区：[1]南通市第三人民医院超声科,南通市226006 [2]南通大学附属医院 [3]National Research Council of Canada

出　　处：《中国超声医学杂志》2020年第3期218-220,共3页Chinese Journal of Ultrasound in Medicine

基　　金：2017年南通市应用基础研究关键技术重点研发(No.MS12017010-6)。

摘　　要：目的运用Faster R-CNN特征计算借助深层CNN架构,分析预标识肝脏肿块超声图像,尝试建立检测器并测试其效能。方法选择肝囊肿及肝癌超声图像为研究对象,收集正常肝脏各切面图像行deep CNN学习,迁移学习后优化预先训练的deep CNN构建更快的R-CNN。将ImageJ软件标识的肿瘤图像作为补丁训练分类器,并通过与基于区域建议的卷积神经网络集成构建检测器,检测器检测样本后自动标识肝脏异常病灶。结果(1)Faster R-CNN较传统检测器检测效率提高;(2)Faster R-CNN预测肝囊肿及肝癌的平均准确率均高于传统HOG-SVM,AlexNet、GoogleNet、ResNet三种CNN预测肝囊肿的准确率差异不显著,而三种方法中ResNet预测肝癌的准确性最佳。deep CNN进行特征转移五次交叉验证后,补丁分类结果中AlexNet、GoogleNet、ResNet预测准确性分别为94.94%、94.14%、98.68%,较传统HOG-SVM分类器准确性87.29%有提高。结论基于deep CNN的Faster R-CNN可高效准确预测肝脏肿瘤超声图像,具有一定的临床及研究价值。Objective To analyze pre-labeled hepatic ultrasound masses using Faster R-CNN deep object detector,and test its effectiveness.Methods We studied ultrasonographic images of hepatic cysts and hcc.ImageJ software was used to label the abnormalities and the labeled patch images were first used to train deep CNN classifiers.The patch classifiers were then integrated with region proposal networks to build faster R-CNN models through transfer learning.The fine-tuned faster R-CNN models are then tested on new ultrasound images for performance evaluation.Results(1)Faster R-CNN is more efficient than traditional detectors.(2)The average accuracy of Faster R-CNN in predicting hepatic cysts and hcc is higher than that of traditional HOG-SVM.The accuracy of AlexNet,Google Net and ResNet in predicting hepatic cysts is not significantly different,while the accuracy of ResNet in predicting hepatocellular carcinoma is significantly higher than AlexNet and GoogleNet.After five cross-validations on feature transfer between deep CNN and SVM,the accuracy of patch classifiers using AlexNet,Google Net and ResNet was 94.94%,94.14%and 98.68%respectively,which was higher than that of traditional HOG-SVM detector(87.29%).Conclusions Faster R-CNN deep detector can effectively and accurately detect liver mass in ultrasound images,and has certain clinical and research value.

关 键 词：DEEP CNN FASTER R-CNN 机器学习 人工智能 超声 肝脏 

分 类 号：R445.1[医药卫生—影像医学与核医学] R735.7[医药卫生—诊断学]