CT Image Detection Of Pulmonary Nodules Based On Deep Convolution Network

It can be seen from the data released by the World Health Organization in 2021 that among all other common types of cancer,lung cancer still has the highest mortality and is still one of the most destructive diseases.The five-year average survival rate is 17.7%.If it is diagnosed and located early,it will increase to 54.4%,but only 15% of cases are diagnosed early in time which can potentially save many lives.Computed tomography(CT)can be used for early detection of pulmonary nodules.The results of lung cancer screening test(LCST)showed that the 5-year mortality of lung cancer screened by CT decreased by about20%.The pulmonary nodule detection system can assist doctors in interpreting diagnostic data.Therefore,it is very important for lung cancer patients to improve the auxiliary detection system of lung nodules to accurately detect lung nodules and effectively reduce the missed detection rate and false positive rate.This study deeply discusses the technical status of lung parenchyma segmentation and lung nodule detection.Taking the deep convolution network and lung nodule CT image as the main research object,it carries out the research on the two parts.(1)Aiming at the main problems existing in lung parenchyma segmentation,this paper proposes an improved double attention mechanism segmentation algorithm based on Ushaped coding structure.The algorithm improves the generalization ability and robustness of network model by appropriately adding residual structure in the coding part,increasing the number of network layers and using double attention mechanism for information fusion.The improved network optimizes the segmentation effect of lung parenchyma,and can accurately find the specific location of lung region,including accurate target boundary.Experiments on LUNA16 and MSD data sets show that the dice similarity coefficient of this method is 95.83%,the performance is better than other methods,and the segmentation effect is better.On the basis of ensuring the robustness,when the convolution network feature learning is good enough,the algorithm is more likely to become a tool to reduce the manual annotation of doctors.(2)For the detection of pulmonary nodules,low sensitivity and difficult recognition of small targets,an improved multi-scale information fusion network based on YOLOv3 is proposed in this paper.Based on the current YOLOv3 backbone network with high recognition accuracy,the multi-scale information fusion network skillfully adds the residual structure,integrates the channel and spatial attention module,adds multi-scale convolution for information fusion,improves the activation function and loss function,and uses data enhancement.Finally,special pulmonary nodules can also be detected by the model.The ablation experimental results on LUNA16 data set show that the modules in the multi-scale information fusion network can work together to better detect positive pulmonary nodules.Compared with the classical model,the sensitivity of the algorithm is 93.4%,and the average CPM score is 0.909.The multi-scale information fusion network can achieve better performance in pulmonary nodule detection,reduce the workload of doctors and improve the diagnosis efficiency.