基于深度迁移学习的肺结节辅助诊断方法  

Pulmonary Nodule Auxiliary Diagnosis Method Based on Deep Transfer Learning

作　　者：张驰名 王庆凤[1] 刘志勤[1] 黄俊[1] 周莹[2] 刘启榆[2] 徐卫云[2] ZHANG Chiming;WANG Qingfeng;LIU Zhiqing;HUANG Jun;ZHOU Ying;LIU Qiyu;XU Weiyun(School of Computer Science and Technology,Southwest University of Science and Technology,Mianyang,Sichuan 621010,China;Mianyang Central Hospital,Mianyang,Sichuan 621010,China)

机构地区：[1]西南科技大学计算机科学与技术学院,四川绵阳621010 [2]绵阳市中心医院,四川绵阳621010

出　　处：《计算机工程》2020年第1期271-278,共8页Computer Engineering

基　　金：四川省军民融合研究院开放基金(2017SCII0219,2017SCII0220);四川省科技创新苗子工程重大项目(19MZGC0123)

摘　　要：在肺癌早期筛查过程中,人工诊断胸部CT扫描图像费时费力,而深度学习网络缺乏足够的医学数据进行训练。为此,提出一种渐进式微调(PFT)策略,将其应用于深度迁移学习网络以辅助诊断肺结节良恶性。利用神经网络在粗粒度的自然图像大数据集中学习特征知识,经重构网络分类层将所学到的特征信息迁移至肺结节的细粒度小数据集。采用PFT策略从全连接分类层开始,逐层释放、微调训练卷积层直至所有网络层,并通过定量分析各层微调后肺结节良恶性分类的AUC值,确定最佳微调深度。此外,采用梯度加权类激活映射图和t-SNE算法为网络预测结果提供相应的视觉支持与解释。在LIDC数据集中的实验结果表明,该方法对肺结节良恶性诊断的准确率可达91.44%,其AUC值为0.9621。In the early screening process of lung cancer,the manual diagnosis of chest CT scan image is time-consuming and laborious.The deep learning network seems like an effective solution,but it lacks sufficient medical data for training.To address this problem,this paper proposes a Progressive Fine-Tuning(PFT)strategy,and applies this strategy to the deep transfer learning network for the auxiliary diagnosis of benign and malignant pulmonary nodules.First,the neural network is used to learn feature knowledge in the large dataset of coarse-grained natural images.Then,the learnt feature information is transferred to the small dataset of the fine-grained pulmonary nodule through the reconstructed network classification layer.From the full-connected classification layer to the convolutional layer,the PFT strategy is adopted to release and fine-tune the layers one by one.Finally,the optimal fine-tuning depth is determined according to the quantitative analysis of AUC values of each layer after fine-tuning.Besides,the Gradient-weighted Class Activation Mapping(Grad-CAM)and t-SNE algorithm are used to provide corresponding visual support and interpretation for network prediction results.Experimental results on the LIDC dataset show that the diagnosis accuracy of benign and malignant pulmonary nodules of the proposed method can reach 91.44%,and its AUC value is 0.9621.

关 键 词：迁移学习 卷积神经网络 医学图像分类 计算机辅助诊断 肺结节诊断 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]