基于区域建议网络的肺结节检测及其去假阳性

如今，基于深度学习的肺结节检测技术不断发展，在辅助医生进行肺结节检查的任务中极大提升了肺结节的检出率和诊断的准确率。本文采用了深度学习技术，提出了一种基于RPN网络结构的肺结节检测方法。针对肺结节的去假阳性阶段，本文将多个分类网络进行了性能对比。本文在Faster R-CNN网络上进行改进，使用Squeeze-and-Excitation（SE）结构以及ResNeXt的残差块构成特征提取模块，再结合UNet++网络结构，输出多个尺度的结果。最后将多尺度结果应用在3D RPN候选检测网络和R-CNN网络上，得到了灵敏度较高，假阳率更低的候选结节检测网络。在去假阳性结节网络阶段，本文用3D DCNN网络对候选肺结节进行假阳性的筛除，有效去除了部分假阳性肺结节，提升了多个FP/scan检查点的灵敏度。本文的网络最终得出灵敏度98.8%@8FPs，Competition Performance Metric（CPM）达到0.879。在去假阳性结节方面，本文验证了3D DCNN网络在几个网络中能够取得最好的效果，达到了15.6%的去假阳率。总的来说，本文的候选结节网络进一步提升了检测的灵敏度，网络模型达到了较好的检测效果。在去假阳性网络方面，得出3D DCNN作为去假阳性网络具有比其它一些网络模型更好的效果。

Improvement of RPN network and research of multiple removing false positive pulmonary nodules networks

The detection technology of pulmonary nodules based on the deep learning method is constantly developing, improving the detection rate and diagnostic accuracy of pulmonary nodules, assisting doctors in the task of pulmonary nodules examination. In this paper, a deep learning method based on RPN network structure was proposed to detect pulmonary nodules, as well as an improved Faster Region-based Convolutional Neural (Faster R-CNN) network was investigated. The performance of multiple classification networks was compared for the false-positive phase of pulmonary nodules. The Squeeze-and-Excitation (SE) structure and ResNeXt residual block were used to form the feature extraction module. Combined with the UNet++ network structure, the multi-scale results were output. Finally, the multi-scale results were applied to the 3D RPN candidate detection network and R-CNN network, then the candidate nodule detection network with higher sensitivity and lower false positive rate was obtained. In the stage of the false-positive nodules task, a 3D DCNN network was used to remove false positive pulmonary nodules for candidate pulmonary nodules, effectively removing some false positive pulmonary nodules and improving the sensitivity of multiple FPs/scan checkpoints. The results show that the sensitivity of the network was 98.8%@8FPs, and the Competition Performance Metric (CPM) was 0.879. In terms of the removal of false-positive nodules, the 3D DCNN network could achieve the best effect in several networks, reaching a false-positive removal rate of 15.6%. It is concluded that the candidate nodule network in this paper further improved the detection sensitivity, and the network model achieved an outstanding detection effect. In false-positive network removal, as a false positive network, 3D DCNN had a better effect than some other network models.