基于Faster R-CNN的肺结节检测改进方法

An improved lung nodule detection method based on Faster R-CNN

为了解决传统的肺结节检测模型检测敏感度低\,假阳性高的问题,提出了一种基于三维卷积神经网络的肺结节检测模型.首先,为了提高运行速度和网络灵活性,采用了单阶段模式,没有假阳性减少阶段;其次,基于上述模型对Faster R-CNN网络进行了改进,创新地将类3D UNet++体系结构作为其区域建议网络的主干,并采用残差块的灵活嵌套模式.3D UNet++网络和残差块都具有多特征融合特性,增强了该模型对特征提取的能力.实验表明:本文改进的方法在基于LUNA16数据集上的假阳性筛查中,其平均敏感度可达到87%,较UNet++网络增长了7.5%;召回率(敏感度)高达95.5%,较VGG16网络增长了7.36%.由此可见,该改进模型能够明显提高检测敏感度和降低假阳性,可为临床应用提供理论参考.

In order to solve the problems of low detection sensitivity and high false positive of the traditional pulmonary nodule detection model,it wa proposed a lung nodule detection model based on 3D convolutional neural network.Firstly,in order to improve the efficiency and flexibility,a single-stage model without false positive reduction stage was adopted.Secondly,based on the model,the Faster R-CNN network was improved by innovative use of the 3 D UNet++-like architecture as the backbone of its regional recommendation network and a flexible nesting mode of residual blocks.Both the 3D UNet++network and the residual block had multi-feature fusion features,which enhanced the model′s ability to extract features.The experiments showed that the improved method achieved an average sensitivity of 87%in the false positive screening based on the LUNA16 data set,which had an increase of 7.5%compared to the UNet++network;when the number of candidate nodules was 48,the recall rate was as high as 95.5%,which had an increase of 7.36%compared to the VGG16 network.This showed that the improved model significantly improved the detection sensitivity and false positive reduction,and would provide a theoretical reference for clinical application.