摘要
针对目前的肺结节检测中存在的个体差异、同病异影、同影异病的问题,提出一种大样本条件下的基于Faster-RCNN的肺结节检测算法,对比研究目前的深度学习模型的适应性,给出一种通用的随着样本数量增加肺结节检测率持续提升的策略。首先搭建深度学习的软硬件环境,设置影像数据接口与Faster-RCNN的网络接口匹配;然后搭建Faster-RCNN的单类分类网络,并对网络结构的参数进行调整优化;最后用包含2000例病人的肺结节数据集,通过不同的卷积神经网络模型(包括ZF和VGG),计算CT图像在各自模型中的特征。对测试结果进行分析评估,分别统计其漏检率、检测准确率,并探讨不同训练数量和数据增广类型对最终检测准确率的影响。最终ZF模型的检测准确率为90.82%,准确率的波动方差为13.30%;VGG模型的检测准确率为87.02%,准确率的波动方差为37.10%。ZF模型的波动方差小,检测精确度高,综合考虑,ZF模型对肺结节的检测效果优于VGG模型的检出效果。所提出的肺结节检测技术具有良好的理论价值和工程应用价值。
Aiming to overcome the problems of individual differences and the homograph in the detection of pulmonary nodules,this paper presented a method of automatic recognition of pulmonary nodules based on Faster-RCNN.By comparing the adaptability of the current in-depth learning model,a general strategy was proposed to continuously improve the detection rate of pulmonary nodules with the increase of the number of samples.First of all,the hardware and software environment for deep learning was built;next,the data interface was set to match with the network interface of Faster-RCNN.Secondly,the single-category classification network of Faster-RCNN was set up and the parameters were adjusted.Thirdly,the pulmonary nodules database containing 2000 patients was utilized to train different feature extraction models(including ZF and VGG models),and the features of CT pictures in different networks were calculated.The test results,missed detection rate and detection accuracy were evaluated.Finally,the influence of different training numbers and data augmentation types on the final detection accuracy was analyzed.The accuracy rate of ZF model was 90.82%,the variance of accuracy rate was 13.30%;the detection accuracy of VGG model was 87.02%,the variance of accuracy rate was 37.10%.Taking into account the balance between the missed detection rate and detection accuracy rate,the ZF model showed small fluctuation variance,a slight low accuracy,and high detection precision.Therefore,the ZF model for pulmonary nodules was better than VGG model.Our proposed lung nodule detection technology has a good theoretical value and engineering application value.
作者
宋尚玲
杨阳
李夏
冯浩
Song Shangling;Yang Yang;Li Xia;Feng Hao(The Second Hospital of Shandong University,Jinan 250033,China;School of Information Science and Engineering,Shandong University,Qingdao 266237,Shandong,China)
出处
《中国生物医学工程学报》
CAS
CSCD
北大核心
2020年第2期129-136,共8页
Chinese Journal of Biomedical Engineering
基金
国家重点研发计划项目(2018YFC0831100,2017YFC0803400)。
