针对传统计算机辅助检测系统中肺结节检测存在大量假阳性的问题,提出一种基于三维卷积神经网络的肺结节识别方法。首先,将传统二维卷积神经网络扩展为三维卷积神经网络,充分挖掘肺结节的三维特征,增强特征的表达能力;其次,将密集连接网络与SENet相结合,在加强特征传递和复用的同时,通过特征重标定自适应学习特征权重;另外,引入focal loss作为网络的分类损失函数,提高对难样本的学习。在LUNA16数据集上的实验结果表明:与当前的主流深度学习算法相比,所提网络模型在平均每组CT图像中假阳个数为1和4时的检出率达到了0.911和0.934,CPM得分为0.891,优于大部分主流算法。

Herein, a method of pulmonary nodule recognition based on a three-dimensional (3D) convolution neural network (CNN) is proposed to overcome the problem of false positives in pulmonary nodule detection by traditional computer aided detection systems. First, a traditional two-dimensional CNN is extended to 3D CNN to fully extract the 3D features of pulmonary nodules and enhance the expressive ability of the features. Second, dense connection network and SENet are combined to enhance feature transfer and reuse, and feature weights are adaptively learned by feature recalibration. In addition, focal loss is introduced as the network classification loss to improve the learning of hard examples. The experimental results on the LUNA16 dataset demonstrate that the proposed network model achieves sensitivities of 0.911 and 0.934 at one and four false positives per scan, respectively, and the competition performance metric is up to 0.891, which is better than that of existing mainstream methods.