Research On The Construction Of Benign And Malignant Pulmonary Nodules Prediction Model Based On Multi Resolution 3D Deep Learning Network

In the analysis of lung computed tomography(CT)scans,the accurate diagnosis of benign and malignant pulmonary nodules is an important research direction of precision medicine,which is of great significance in practical clinical applications.Traditional methods of manual diagnosis to diagnose the benign and malignant pulmonary nodules not only depend on doctors’ diagnostic level,but also have a strong subjectivity,and will bring huge workloads.However,deep learning methods can not only deeply explore the complex implicit representation of image data,but also further analyze and process those features to make classification or prediction tasks more intelligent.Therefore,a Multi-Resolution 3D Dual Path Squeeze and Excitation deep learning network model is constructed to predict the benign and malignant of pulmonary nodules in lung CT images in this thesis.In the processing of pulmonary nodule image data,in order to make full use of the irregular information of the spatial shape of the pulmonary nodules,meet the requirements of the deep learning network for the fixed input data dimension,and to benefit the deep learning network to learn the characteristics of pulmonary nodules.In this thesis,a complete 3D pulmonary nodule is extracted from the CT image data as an input to the deep learning network,and a 3D multi-resolution data processing method is proposed.This method can effectively deal with problems such as boundary loss and noise caused by different diameters of lung nodules.In the construction of 3D deep learning network model,firstly,the dual path network(DPN)is transformed into a 3D structure;secondly,by embedding the squeeze and excitation(SE)unit into the 3D DPN to constitute a 3D dual path squeeze and excitation network.The network can re-calibrate the weights of the feature channel,and can effectively describe the importance of different feature channels to the network output.Then,a pile of theoretical and verification test analysis of the 3D dual path squeeze and excitation network model is carried out,and the network structure is rationally adjusted and optimized.Finally,a multi-resolution 3D dual path squeeze and excitation network model is constructed based on the multi-resolution method.This multi-resolution 3D deep learning network model can not only reuse low-order features of pulmonary nodule images but also continuously generate new high-order combined features with superior prediction performance.In order to verify the validity of the constructed multi-resolution 3D deep learning network model for predicting benign and malignant pulmonary nodules,the classification accuracy(ACC)and area under curve(AUC)are used as evaluation criteria in this thesis.The LIDC-IDRI dataset and the high-resolution thim-section CT dataset constructed by the team are tested,and an ACC of 87.96%,AUC of 93.11%and an ACC of 86.87%,AUC of 90.36%are obtained,respectively.The experimental results show that the proposed multi-resolution 3D deep learning network model can not only maintain excellent performance when using the largest number of pulmonary nodules but also effectively cope with pulmonary nodules with complex spatial structure compared with other state-of-the-art benign and malignant pulmonary nodules prediction methods.At the same time,a good cross-dataset scalability is demonstrated on the high-resolution thin-section CT dataset,which verifies the validity and robustness of the model.