多输入卷积神经网络肺结节检测方法研究

Study on Detection Method of Pulmonary Nodules with Multiple Input Convolution Neural Network

针对传统计算机辅助诊断系统中肺部结节检出过程复杂,检出结果依赖于分类前期每个步骤的性能,以及存在假阳性率高的问题,提出了一种基于卷积神经网络的端到端的肺结节检测方法。该方法首先使用大量带标签的肺结节数据对构建的多输入卷积神经网络进行训练,实现从原始数据到语义标签的有监督学习。然后采用快速边缘检测方法和二维高斯概率密度函数构建候选区域模板,从待检测CT序列中获取候选区域并将其作为多输入卷积神经网络的输入数据。最后采用判定阈值实现疑似肺结节区域标注,同时在相邻的CT影像中进行重点检测。在LIDC-IDRI数据集上的大量实验结果表明,所提方法在肺部CT影像中对微、小结节的检出率较高;同时,重点检测模板能够小幅降低微、小结节检测的假阳率。

In view of that the detection process of pulmonary nodules in traditional computer-aided diagnosis system is complex,the detection results depend on the performance of each step in the early stage of classification,and there is a problem of high false positive rate,this paper presented an end-to-end detection method of pulmonary nodules based on convolution neural network.First,it uses a large number of tagged pulmonary nodule data to input the newly constructed multi-input convolution neural network for training,realizing the supervised learning from the raw data to the semantic label.Then,it uses the fast edge detection method and the two-dimensional Gaussian probability density function to construct the candidate region template,and the obtained candidate region from the CT sequence is used as the input data of the multi-input convolution neural network.Finally,it uses a diagnostic threshold to annotate the suspected pulmonary nodule region,which will be emphatically checked in the next frame.A large number of experimental results on LIDC-IDRI data set show that the detection rate of the small nodules in the lung CT image with the proposed method is high,and the candidate region template with key monitor can slightly reduce the false positive rate of the small nodules detection.