Abstract: 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.

Key words: Computer-aided diagnosis,Convolution neural network,Pulmonary nodule detection,LIDC-IDRI dataset preprocessing