Detection Of Low-dose CT Pulmonary Nodule Based On Convolutional Neural Network

Lung cancer is one of the fastest growing tumor diseases in the world with morbidity and mortality.70% to 80% of lung cancer patients are found to be in the middle and late stages.At this time,the lesion has been unable to be surgically removed.,although it can be treated by radiotherapy and various chemotherapy regimens,but the prognosis of advanced lung cancer is extremely poor,with an average survival time of less than 12 months.Therefore,early detection of lung lesions to improve the rate of complete resection(ROI resection)is of great significance for improving the survival rate of lung cancer patients.The application of the Computer Aided Diagnosis System(CADe)greatly improved the diagnostic rate of early nodule.However,since the existing traditional CADe methods are mostly limited to the influence of subjective factors such as morphological features,the accuracy rate needs to be improved and the false positives are high,while the current deep learning detection method is more basic,although the false positive rate is effectively reduced.However,the accuracy rate will also decrease.In order to solve this problem,this paper studies the low-dose CT image lung nodule detection method based on convolutional neural network.We solve the problem by combining the improved whole convolutional neural network(CU-net)and the cyclic 3D Faster-Rcnn(3D CFaster-Rcnn)lung nodule detection method.Firstly,CU-net is used to perform candidate region detection on CT images,and then suspected nodule regions of the images are quickly located without changing size of the images.The candidate area three-dimensional pixel block is extracted and trained in the 3D CFaster-Rcnn model by pseudo-region coordinate calculation,as well,the false positive removal is operated.It is obtained that the nodule recall rate of the candidate region detection step is 98.5%.After the model optimization to false positive treatment,the accuracy rate achieves 92.6% with false positive rate being 1.65.Compared with other methods,the model has higher accuracy when the false positive is in low level.After pre-processing with the Data Science Bowl 2017 data set,the results showed that the nodule recall rate of the candidate region detection step was 98.5%.After the false positive treatment,that is,the model optimization,92.6% was obtained when the false positive rate was 1.65.The accuracy rate.Compared with other methods,the model has higher accuracy when the false positive is lower,and has higher application value.Through the experimental process and results,it is shown that deep learning has a good development trend and practical significance in the detection of pulmonary nodules.