一种改进三维卷积模型的假阳性肺部结节筛除方法

False positive pulmonary nodule screening method with improved 3D convolution model

针对在肺结节候选区域中假阳性率过高的问题,采用3D CNN对假阳性结节进行筛除,引入空洞卷积代替部分池化操作,在增大感受野的同时尽可能多地保留特征信息,解决了在传统卷积和池化层中对形状不规则且尺寸较小的肺结节无法高效能地收集到肺结节的像素点问题。在公开的LUNA16数据集中,AUC的值可以达到0.967,说明该模型对正负样本的分类能力较好;97%的特异值和88%的敏感度,表明了该模型可以有效地避免误检且漏检的可能性也较小。实验表明提出的三维卷积神经网络适用于降低肺结节检测中的假阳性率。

In view of the too high false positive rate(FPR)in the candidate region of pulmonary nodules,3D CNN(convolutional neural network)is used to screen the false positive nodules,the cavity convolution is introduced instead of partial pooling operation,which keeps as much feature information as possible while increasing the receptive field,so that the problem of collecting pixels of pulmonary nodules with irregular shape and small size in the traditional convolution and pool layer is solved.The AUC value can reach 0.967 in the open LUNA16 data set,which indicates that the model has better classification ability for positive and negative samples.The specificity of 97%and the sensitivity of 88%indicate that the model can effectively avoid false detection and its possibility of missed detection is small.The experiments show that the proposed 3D CNN is suitable for reducing the FPR in pulmonary nodule detection.