基于改进Faster R-CNN和3D DCNN的肺结节检测算法

针对传统肺结节检测准确率低，且存在假阳性高的问题，提出了一种改进Faster R-CNN网络检测候选结节，以及改进的3D DCNN网络去除假阳性的算法。考虑到结节的形状大小等因素，在Faster R-CNN上更改锚点数量和尺寸检测结节的鲁棒性，并在特征提取器的最后一层添加一个反卷积层，在网络特征图上根据结节尺寸添加小型滑动网络以使网络自适应生成感兴趣区域，获取候选结节。为了去除假阳性结节，在2D DCNN网络基础上调整卷积核参数，引入时间维度生成3D DCNN，并利用Adam算法调整网络学习率更改网络权重参数，采用数据增强策略进一步提取结节的全局特征。LIDC-IDRI数据集上的实验结果表明，所提出的算法平均检测准确率达到97.71%，同时降低了误诊率和漏诊率。

Pulmonary nodule detection based on improved Faster R-CNN and 3D DCNN

Aiming at the low accuracy of traditional lung nodule detection and the high false positive, this paper proposed an improved Faster R-CNN network and improved 3D DCNN to detect candidate nodules and remove false positives. Considering the shape and size of the nodule, etc., the method changed the anchor point number and size to detect the robustness of the nodule on Faster R-CNN, and added a deconvolution layer in the last layer of the feature extractor. In addition, according to the size of the nodule, the method added a small sliding network to enable the network to adaptively generate the region of interest to obtain candidate nodules. In order to remove false positive nodules, it adjusted the convolution kernel parameters based on the 2D DCNN, used the time dimension to generate the 3D DCNN, and applyied the Adam algorithm to adjust the network learning rate to change the network weight. The enhancement strategy further extracted the global features of the nodule. The experimental results on the LIDC-IDRI dataset show that the proposed algorithm has an average detection accuracy of 97.71%, and reduces the rate of misdiagnosis and missed diagnosis.