基于深度卷积神经网络的两阶段肺结节检测

Two-stage pulmonary nodule detection based on deep convolutional neural network

针对传统肺结节检测中存在灵敏度低、假阳性高、小结节难检测的问题,提出一种基于深度卷积神经网络的两阶段肺结节检测框架。第一阶段使用特征金字塔子网提取肺部影像的多层次特征,引入多尺度区域建议子网用于在高灵敏度下检测出所有的候选结节;第二阶段设计级联卷积神经网络模型减少假阳性,通过保留分类错误样本用于重新训练模型,将多个模型结果进行投票选出最终分类结果。LUNA16数据集上的实验结果表明,所提框架灵敏度达到95.9%,检测效果优于其它算法,能够有效实现肺结节的准确检测。

Aiming at the problems of low sensitivity,high false positives,and difficult detection of small nodules in traditional pulmonary nodule detection,a two-stage pulmonary nodule detection framework based on deep convolutional neural networks was proposed.In the first stage,the feature pyramid subnet was used to extract the multi-level features of the pulmonary image,and the multi-scale region proposal subnet was introduced to detect all candidate nodules at high sensitivity.In the second stage,a cascade convolutional neural network model was designed to reduce false positives,the classification error samples were retained to retrain the model,and multiple model results were voted to select the final classification result.Experimental results on the LUNA16 dataset show that the sensitivity of the proposed framework reaches 95.9%,and the detection effect is better than other algorithms,which can effectively achieve accurate detection of pulmonary nodules.