Research On Accurate Detection Method Of Pulmonary Nodules Based On Low Dose CT Images

Lung cancer has the highest incidence and mortality of malignant tumors in the world.Early detection,early diagnosis and early treatment are the keys to saving the lives of patients with lung cancer.The early clinical manifestation of lung cancer is pulmonary nodule.Pulmonary nodule screening based on low dose Computed Tomography(CT)is an important method to detect early lung cancer.However,it is a great challenge to rely only on radiologists to find tiny pulmonary nodules from hundreds of chest CT images.Therefore,in order to meet the clinical needs of rapid and automatic detection of the exact position of pulmonary nodules in CT images,this thesis proposes a method for accurate detection of pulmonary nodules based on low dose CT images.In order to solve the problem of low sensitivity caused by missed detection of small nodules and reduce the false positive rate,this thesis proposes an accurate detection method combining two-dimensional and three-dimensional information.Firstly,the improved Faster R-CNN detection model is designed to realize the preliminary automatic accurate screening of pulmonary nodules,and then the 3D Inception CNN network is designed to reduce the false positive rate.The main contents of this thesis are as follows:(1)In order to solve the problem of low sensitivity caused by missed detection of small nodules,this thesis proposes an improved Faster R-CNN pulmonary nodule detection model.Firstly,we use feature fusion technology to concatenate the shallow and deep features of the feature extraction network to enhance the feature extraction capability of the model.Then we optimize the region proposal network(RPN),use three parallel convolution kernels of different sizes to replace the original 3×3convolution to generate better candidate regions.Finally,the Region of Interest(ROI)Pooling layer is optimized to enhance the detection ability of small nodules.The experimental results show that the proposed method has a better detection effect on small nodules which are easy to miss.The sensitivity of the improved model is94.63%,and the false positive number of each CT scan is 14.6.The detectionsensitivity is 7.94% higher than that of the Faster R-CNN model,and the false positive number of each CT scan is reduced by 8.1.(2)In order to reduce the false positive rate,this thesis designs a 3D Inception CNN Network with Residual Conv Block and Spatial Reduction Block.The Spatial Reduction Block adopts a structure in which the convolution layer and the pooling layer are operated in parallel first,and finally combined.It not only achieves a "gentle" reduction in the size of the feature map,but also performs multi-scale and multi-dimensional feature extraction and fusion operations on the upper layer features to retain the integrity of the information,which is conducive to the improvement of network classification accuracy.The Residual Conv Block can be regarded as a combination of the residual unit and the inception module,which can simplify learning difficulty,and also contribute to the improvement of network performance.The combination of these two modules can better extract network features and improve network performance.Experiments show that the model designed in this thesis can achieve a more accurate classification effect with a relatively small number of parameters.The final classification accuracy is 95.2%,the sensitivity is 95.7%,and the specificity is 94.6%,which has certain practical significance.The two-dimensional and three-dimensional combined pulmonary nodule accurate detection method proposed in this thesis not only realizes the automatic and accurate detection of candidate nodule positions on two-dimensional CT images,but also makes full use of three-dimensional volume data to reduce the false positive rate.Compared with the detection algorithm using only two-dimensional information,the method proposed in this thesis improves the detection accuracy and reduces the false positive rate.Compared with the detection algorithm using only three-dimensional information,the method proposed in this thesis reduces the amount of calculation to a certain extent,and has certain theoretical and practical value.