Detection And Recognition Of Pulmonary Nodules In CT Images Based On Deep Learning

The detection and recognition of lung nodules is an important part of the lung computer-aided diagnosis system.Aiming at the problems of high false positives,high false alarm rate and variable size of nodules in the process of lung nodule detection and identification,a new lung nodule detection and recognition model was established by using deep learning methods.The main research content of the thesis includes the following three aspects:(1)The research foundation of the thesis: First,the development history and mainstream trends of medical image processing technology in lung CT image processing are introduced,and then the method of constructing and preprocessing lung CT image data sets is described,and finally lung nodules are given.Mainstream evaluation indicators for detection and identification.(2)A method for detecting lung nodules based on 3D NAS-FPN is proposed.First,considering the characteristics of the 3D characteristics of lung nodules and the variety of nodule scale changes,the neural network search architecture is used to search for the expandable FPN structure,and then the model is optimized through the experimental results,and finally the One-Stage and Two-Stage algorithms are selected.The better-performing YOLOv4 and Faster-RCNN algorithms are used as comparative experiments to verify the effectiveness of our proposed model in detection.(3)A 3D Attention DPN model for benign and malignant pulmonary nodules recognition based on attention mechanism is proposed.Using the compactness and high efficiency of the 3D DPN network,combined with the attention mechanism to adjust the weight of this network,a 3D Attention DPN network based on the attention mechanism is designed to extract the nodule features,combined with GBM(Gradient Boosting Machine)The algorithm classifies benign and malignant nodules,and the entire network can quickly reach the optimal state through adjustment.Finally,several commonly used recognition algorithms are selected for comparison.The experimental results show that our proposed model is better than the comparison model.