基于改进YOLO V3的肺结节检测方法  

Pulmonary nodule detection based on improved YOLO V3

作　　者：王乾梁 石宏理[1,2] WANG Qianliang;SHI Hongli(School of Biomedical Engineering,Capital Medical University,Beijing 100069,China;Beijing Key Laboratory of Fundamental Research on Biomechanics in Clinical Application,Capital Medical University,Beijing 100069,China)

机构地区：[1]首都医科大学生物医学工程学院,北京100069 [2]首都医科大学临床生物力学应用基础研究北京市重点实验室,北京100069

出　　处：《中国医学物理学杂志》2021年第9期1179-1184,共6页Chinese Journal of Medical Physics

基　　金：北京市自然科学基金(7142022).

摘　　要：针对肺结节占CT图像比例小、形状不规则及直接应用YOLO V3算法进行肺结节检测效果不佳的问题,提出基于改进YOLO V3的肺结节检测方法。首先进行重采样和肺实质分割等预处理操作。然后修改YOLO V3的基础网络结构,调整骨干网络和检测网络的结构单元数量;使用Mish激活函数替换Leaky ReLU激活函数,引入含有空洞卷积的感受野模块层;修改损失函数。最后使用改进的YOLO V3方法进行肺结节检测,完成对比实验。在LIDC-IDRI数据集上得到了88.89%的准确率和94.73%的高敏感度,实验结果表明该方法能够有效检测肺结节。In view of small proportion of pulmonary nodules in CT images,irregular shapes of pulmonary nodules and unsatisfactory results obtained by the direct application of YOLO V3 algorithm for pulmonary nodule detection,a pulmonary nodule detection method based on improved YOLOV3 is proposed in the study.Preprocessing such as resampling and parenchymal segmentation is carried out.Then the basic network structure of YOLO V3 is modified,and the number of structural units of the backbone network and the detection network is adjusted.Leaky ReLU activation function is replaced by Mish activation function,and receptive field block layers with dilated convolutions are added.Moreover,the loss function is modified.Finally,the improved YOLO V3 is used to detect pulmonary nodules,and the comparative experiment is completed.The proposed method is tested on LIDC-IDRI data set,and the results show that the improved YOLO V3 achieves an accuracy of 88.89%and a sensitivity of 94.73%,indicating that the proposed method can effectively detect pulmonary nodules.

关 键 词：深度学习 肺结节 YOLO V3 卷积神经网络 

分 类 号：R318[医药卫生—生物医学工程] TP391.41[医药卫生—基础医学]