Research On Automatic Detection Algorithm Of Pulmonary Nodules Based On Deep Learning And Feature Mixing

Lung cancer is a malignant tumor disease with the highest morbidity and mortality in the world,and its early diagnosis can effectively improve the survival rate of patients.Lung nodule is an early manifestation of lung cancer,which is often detected by CT.However,with the explosive growth of medical imaging data,it brings great work pressure to imaging doctors,which leads to doctors' false detection and missed detection due to their subjective will.With the development of deep learning technology,it gradually plays an irreplaceable role in medical image processing,which can effectively analyze images and assist doctors in diagnosis,and has important practical application value.To improve the working efficiency of imaging doctors and reduce the risk of misdiagnosis and missed detection of pulmonary nodules.In this paper,following the software engineering specification,the waterfall model is used for software development.Based on the analysis of users' functional requirements and performance requirements,the outline design and detailed design of the system are carried out.Python,Java and other programming languages are used to build an automatic lung nodule detection algorithm based on deep learning and feature mixing to complete the development and testing of the auxiliary lung nodule detection system.The system takes the auxiliary detection of pulmonary nodules as the core and the pulmonary nodules in CT images as the research object.Aiming at the problems of high false positive,low detection rate and difficult localization of small-scale nodules in current pulmonary nodule detection algorithms,an automatic detection algorithm based on deep learning and feature mixing is proposed.Firstly,in order to improve the quality of the original image,HU threshold processing,region growing algorithm and morphological operation are used to process the original input image.Secondly,R-FCN is selected as the backbone network and improved,so that it can efficiently complete the task of pulmonary nodule detection.DenseNet is selected as the feature extraction network to enhance the feature reuse ability of the network layer,and K-means algorithm is used to cluster the boundary boxes of pulmonary nodules.The results are used to modify the Anchor-related settings in RPN network,which makes the network suitable for pulmonary nodule detection tasks.To solve the problem of rough localization of small-scale nodules,FPN is introduced into RPN network to make full use of the semantic features and location features of pulmonary nodules,so as to achieve the purpose of accurate localization and prediction of pulmonary nodules.Finally,Focal loss function is introduced in the training process to solve the imbalance between positive and negative samples in the model training process.Experimental results show that the accuracy of lung nodule detection using the proposed algorithm can reach 94.94%.The system test proves that the auxiliary detection system for pulmonary nodules has achieved various functions,which can effectively assist doctors in intelligent diagnosis,save manpower and time investment in CT image reading,improve the diagnostic efficiency of imaging doctors for pulmonary nodules and reduce the risk of missed diagnosis.In addition,the realization of the system is helpful for patients to know their own situation in time,which has good practical significance.