Research On CT Image Pulmonary Nodule Detection Based On Convolutional Neural Network

Lung cancer is the cancer with the highest morbidity and mortality in the world,and its early diagnosis and treatment can effectively improve the survival rate of patients.Pulmonary nodules are the main manifestation of lung cancer in the early stage,so it is of great significance to locate and discriminate lung nodules.At present,the popularity of CT imaging equipment has made it the main imaging method for early screening of lung cancer,but the large number of CT images has greatly increased the burden of radiologists.The computer-aided diagnosis of pulmonary nodules technology can improve diagnosis efficiency and reduce the subjectivity and instability of manual image reading by locating suspicious lesion areas for physicians’ reference.To improve the performance of auxiliary diagnosis,the advantages and disadvantages of existing medical image processing methods are analyzed.The lung parenchyma segmentation method and lung nodule detection method are studied.The overall process is to first segment the lung parenchyma to roughly extract the lesion area,and then detect the lung nodules to finely locate the lesion.The main work is as follows:(1)The method of lung parenchymal segmentation based on improved U-net is studied.Aiming at the problems of low segmentation accuracy and slow convergence speed of existing lung parenchymal segmentation methods,a segmentation method consisting of two stages of presegmentation and fine segmentation is proposed.In the pre-segmentation stage,the K-means clustering method and morphological operations are used to binarize the CT slices,and the convex hull scanning method is used to repair the missing features at the target boundary to realize the preliminary positioning and correction of the lung parenchyma.In the fine segmentation stage,the Sobel convolutional layer is used to strengthen the high-pass filtering of the edge area,and the RADUnet(Residual Atrous Dense U-net)network model is designed based on the U-net structure.By including the residual structure of the pre-activated module,the atrous convolution and the dense connection are improved to achieve the refined segmentation result.By using a segmentation method that combines traditional image segmentation methods with deep learning strategies,the average Dice similarity coefficient on the experimental test set is 98.49%,and the Jaccard similarity coefficient is97.02%,and a good segmentation effect and performance are obtained.(2)The method of lung nodule detection based on an improved feature pyramid network is studied.Aiming at the characteristics of pulmonary nodules in size diversity and location randomness,a detection method based on the Dual-Path Fusion Feature Pyramid Network is proposed.First,a bottom-up feature fusion path on the basis of the feature pyramid network is used to improve the spreading ability of shallow features.Then,the three-dimensional context attention mechanism is introduced when fusing features to improve the flexibility of the model.Next,a residual structure combined with depth separable convolution is designed to make the model tend to be lightweight.Finally,the candidate region generation module of Faster R-CNN is extended to three dimensions to predict the multi-scale fusion feature map.The average sensitivity of the method reached 96.12%,and the CPM value was 86.34%.By comparing quantitative and qualitative experiments,the good performance of the method in terms of nodule recall and precision is verified.(3)An auxiliary detection system for lung nodules is designed and implemented.First,demand analysis from the perspective of function and performance is conducted.Then,based on the modular idea,seven functional modules including login,consultation,data interaction,lung parenchymal segmentation,lung nodule detection,model management,and medical record management are designed.Next,the addressing method is used to access the key image data,and the database core table structure is designed.Finally,the module development and human-computer interaction interface of the system are realized based on Python and QT platform.The proposed algorithm model and a variety of state-of-the-art models are integrated into the system to assist physicians in completing the entire process of patient diagnosis,thereby improving the diagnostic efficiency and accuracy of the physician.