Pulmonary Nodule Detection Based On Manifold Learning And Convolutional Neural Network

Lung cancer has now become the "number one killer" in cancer in China,and early detection of early treatment is extremely important.The pulmonary nodule detection system is a computer-aided diagnosis system that can help doctors make better diagnosis,improve the accuracy of diagnosis,and effectively reduce the burden on doctors.However,the traditional pulmonary nodule detection system has the disadvantage that the feature extraction is not resistant to noise and the generalization ability is poor.To this end,it is necessary to conduct an in-depth study on the optimization of the pulmonary nodule detection system.Convolutional neural networks are widely used in image detection and segmentation tasks because of their excellent feature extraction and classification capabilities.U-net convolutional neural networks are an efficient method for lung nodule detection,but they have the disadvantage of over-fitting.As a feature extraction algorithm,the local linear embedding algorithm in manifold learning has the disadvantages of poor anti-noise ability and insufficient processing of large curvature data.To this end,the shortcomings of the above two methods are insufficient.Firstly,the U-net convolutional neural network and the local linear embedding algorithm are improved,and then the improved method is used for the deep lung nodule detection research application.The main work and innovations are as follows:(1)In view of the shortcomings of U-net convolutional neural network with poor generalization ability and easy over-fitting,the whole convolutional lung nodule segmentation network(FCLNSN)is studied,firstly using the parameterized Re LU activation function to prevent "gradient" in the convolutional layer."Saturation" phenomenon;reintroduction of random inactivation layer,with reasonable He parameter initialization scheme to prevent over-fitting of network model;and designing network loss function according to input and output characteristics;finally adopting more advanced optimizer Nadam to accelerate model training.The experimental results show that FCLNSN has a higher detection recall than U-net.(2)For the local linear embedding algorithm,it is difficult to effectively deal with noise,large curvature and sparse sampling data,and research on the local linear embedding algorithm(IRWLLE)for improving the reconstruction weight.First,reconstruct and define the reconstruction weight in the LLE,that is,the ratio of the geodesic distance to the Euclidean distance is defined as the structural weight in the neighborhood of a sample.Then,the ratio of the geodesic distance to the median geodesic distance is defined as the distance weight,and the product of the structural weight and the distance weight is used as the reconstruction weight,so that the structure and distance information of the manifold are organic.Combination of.The experimental results show that the IRWLLE algorithm has better robustness than the original LLE algorithm,the ability to process sparsely sampled data,and the ability to process large curvature data,which greatly improves the recognition rate.(3)Study on the detection of pulmonary nodules with FCLNSN and IRWLLE.In order to further improve the precision of the pulmonary nodule detection system,the IRWLLE algorithm was combined with the FCLNSN network.First,the FCLNSN was used to segment the candidate lung nodules,and then IRWLLE was used for feature extraction and false positive screening.The experimental results show that the algorithm framework of FCLNSN+IRWLLE can obtain higher precision and better screening false positive nodules.