Detection Of Pulmonary Nodules Based On Improved Faster R-CNN

Lung cancer is a malignancy with the highest morbidity and mortality rate in both China and the world.Its early manifestation is a circular or irregular nodule in the lung with a diameter of no more than 30 mm.Early diagnosis and treatment of lung cancer is the most effective means to reduce its mortality.The detection of pulmonary nodules is mainly achieved by CT examination,researches shows that,low dose CT can effectively reduce the mortality of lung cancer patients.Therefore,low dose CT is gradually becoming the mainstream tool for pulmonary nodule detection.Deep learning is very popular in medical image field,Faster R-CNN is a target detection algorithm based on deep learning.In view of this,this paper is designed to study the detection algorithm of pulmonary nodules based on improved Faster R-CNN under low-dose two-dimensional CT chest image.The main work of this paper is as follows:（1）Detection of pulmonary nodules based on improved Faster R-CNNIn order to rapidly find the exact position of pulmonary nodules in hundreds of two-dimensional CT chest images,a new detection method of pulmonary nodules based on improved Faster R-CNN was proposed.Firstly,the data enhancement technology was used to enlarge the dataset and then the dataset was manually annotated.Secondly,the image was sent to the VGG16 network added with a layer of deconvolution to extract the shared convolution feature.The shared convolution feature is then sent into the Region Proposal NETWORK（RPN）to output the candidate nodule region.Finally,according to the position and background information of the candidate nodule region in the original image,the features of the corresponding candidate region were extracted on the shared convolution feature,and these features of the candidate regions were sent into the pooled layer of the Region Of Interest（ROI）to obtain a map of the fixed-size candidate region features,Finally,the ROI feature vector was obtained through the whole connected layer,and the ROI feature vector was fed into the Softmax classifier and the boundary regressor,and the classification score of the candidate node region and its boundary box after the regression were given out,which was the final chest image detection result.In view of the complicated background of chest image,the large range of variation of detected objects and the size of the lung nodule that is smaller than the natural object,we have designed a smaller anchor box to adapt to the variation of pulmonary nodule size.In order to obtain a better and more accurate description of the characteristics of pulmonary nodules,we added a deconvolution layer whose size of kernel,stride,filling and kernel number was respectively 4,4,2 and 512 behind the last layer of the VGG16 network conv5₃,which restores a degree of feature with finer grit than the original feature map.The experimental results show that,the average detection accuracy of the improved model is 6.9% higher than that of the original Faster R-CNN model.The improved model can detect the solitary pulmonary nodules and pulmonary nodules adhered to the pulmonary wall as well as small nodules,and it has a certain clinical significance for early screening of lung cancer.However,the improved model has the potential to further improve the accuracy of the detection,and it can be further optimized to reduce false positive rate,so as to increase the accuracy of the detection of pulmonary nodules.（2）Optimization of improved Faster R-CNNFor the improved Faster R-CNN model,the following optimization strategy is chosen: Choose the optimum Dropout value and the best Batch Size.After setting up these two optimum parameters,a difficult negative sample mining technology was added to the improved Faster R-CNN network to improve its ability to detect the negative samples.Dropout can reduce the overfitting of the network,and enhance the generalization ability of the network,and it takes many times of parameter adjustment to determine the best Dropout value.The selection of Batch Size value influences the training speed and precision of the model.The difficult negative sample mining is used to improve the performance of classifier module in classifying the difficult negative sample in the detection system（highly similar to the nodule,but not the same,such as blood vessel,trachea and bronchus.）,it was added to the improved Faster R-CNN network to reduce false positive rate,so as to increase the accuracy of the detection of model.Through the experimental verification and analysis,the accuracy of the optimization model is proved to be better than that of the improved Faster R-CNN model,it not only reduces the false positive rate,but also detects the pulmonary nodules quickly and accurately from hundreds of chest two-dimensional CT images,which lays a good foundation for the early diagnosis and treatment of pulmonary nodules.