改进的NASNet算法在乳腺超声诊断中的应用研究  

Application of Improved NASNet Algorithm in Breast Ultrasound Diagnosis

作　　者：易三莉[1,2] 佘芙蓉 杨雪莲 陈东 罗晓茂 Yi Sanli;She Furong;Yang Xuelian;Chen Dong;Luo Xiaomao(School of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650504,China;Key Laboratory of Computer Technology Application of Yunnan Province,Kunming 650504,China;Department of Ultrasound Medicine,Yunnan Cancer Hospital,Kunming 650118,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,昆明650504 [2]云南省计算机技术应用重点实验室,昆明650504 [3]云南省肿瘤医院超声医学科,昆明650118

出　　处：《中国生物医学工程学报》2022年第6期672-679,共8页Chinese Journal of Biomedical Engineering

基　　金：国家自然科学基金(22174057);云南省教育厅项目(2020J0052)。

摘　　要：超声图像在乳腺疾病的临床诊断中具有重要意义,但是乳腺超声图像分辨率低、样本量小,同时对于NASNet算法虽然适用于小样本数据但参数量大难以训练的问题。本研究提出一种改进的NASNet分类算法来检测乳腺肿块的良恶性。首先采用迁移学习技术将NASNet在ImageNet上预训练,将学习到的特征直接用于乳腺超声图像上肿块的良恶性识别,节省计算成本的同时提高精度;然后,为增强网络对超声图像特征的提取能力并使网络轻量化,在NASNet深层融入深度可分离卷积,构造出更深更宽的网络;最后,为了增强与疾病更加相关的特征权重,进一步增强高阶特征信息的提取能力,加入SE模块来筛选超声图像中占比较重的通道特征。为验证该算法,基于本地医院数据集以及公共数据集实验,其中本地医院数据共计1350张超声图像,两个公共数据集共计895张超声图像。采用五折交叉验证的训练方法,并将该算法与目前广泛应用的分类算法进行比较。基于本地医院数据实验的Acc、Sen、F1均为97.52%;公共数据集作为训练集和验证集,本地医院数据集作为测试集的实验的Acc、Sen、F1分别为96.31%、96.31%、96.39%;基于本地医院数据和公共数据的混合数据实验的Acc、Sen、F1均为98.27%。所提算法较其他算法具有优越性,证实了该算法更适用于小样本乳腺超声图像的肿块良恶性分类。Ultrasound image is of great significance in clinical diagnosis of breast diseases,however,the resolution of the breast ultrasound image is low,and the sample size is small.Although NASNet is suitable for small sample data,it requires many parameters that makes it difficult to train.This paper proposed an improved NASNet classification algorithm to distinguish the benign and malignant breast masses.Firstly,NASNet is pre-trained on Imagenet by transfer learning technology,and the learned features were directly used for benign and malignant tumor recognition on breast ultrasound images,which saved the cost of calculation and improved the accuracy of the calculation.Then,to enhance the ability of the network to extract ultrasonic image features and make the network lightweight,we deeply integrated deep separable convolution into NASNet to construct a large-scale network.Finally,to enhance feature weights that are more relevant to the disease and further enhance the extraction ability of high-order feature information,we added an SE module to screen the channel features that account for more weight in ultrasonic images.To verify the algorithm,we used the training method of 5-fold cross-validations based on the experiments of local hospital data sets and public data sets and compare the algorithm with the widely used classification algorithm.There were 1350 ultrasound images in the local hospital datasets and 895 ultrasound images in the two public datasets.Based on the data of local hospitals,Acc,Sen,and F1 were 97.52%.The Acc,Sen,and F1 of experiments with public data set as training set and verification set and local hospital data set as test set were 96.31%,96.31%,and 96.39%respectively.The Acc,Sen,and F1 of the mixed data experiment based on local hospital data and public data were 98.27%.The results showed that the improved algorithm had advantages over other algorithms and was more suitable for the classification of benign and malignant tumors with a small amount of breast ultrasound images.

关 键 词：乳腺肿块 NASNet 深度可分离卷积 SENet 迁移学习 

分 类 号：R318[医药卫生—生物医学工程]