Research On Automatic Detection Algorithm Of Pulmonary Nodules Based On Multi-scale Fusion Convolutional Neural Network

Globally,the death rate of lung cancer ranks first among all cancers.Pulmonary nodules are an early symptom of lung cancer.Accurate detection and treatment of lung nodules is of great significance for reducing the mortality of lung cancer patients.At present,the screening and diagnosis of lung nodules are mainly achieved through computer tomography(CT)images.However,the size of lung nodules in CT images is not fixed,the shape is changeable,and it is easy to confuse and adhere to the blood vessels,organs and other tissues in the lung.Therefore,it is a very difficult task to accurately detect lung nodules.With the development and application of deep learning in medical images,it has become possible for artificial intelligence technology to assist doctors in quickly diagnosing lung nodules.In this paper,aiming at the problems of incomplete feature extraction of pulmonary nodules in pulmonary nodule detection and pulmonary nodule detection algorithms are easily affected by the size of pulmonary nodules,an automatic detection algorithm of pulmonary nodules based on multi-scale convolution neural network is proposed,and the dense connection strategy is adopted to further improve the detection accuracy of micro pulmonary nodules.The main work is as follows1.Automatic lung nodule detection algorithm based on multi-scale convolutional neural network.Because of the difference in size and shape of pulmonary nodules in CT images,and the similarity with pulmonary vessels in gray scale,it is very difficult to accurately and comprehensively extract the features of pulmonary nodules,which leads to the problem of multiple and missed detection of pulmonary nodules in clinic.This paper proposes an automatic detection algorithm of pulmonary nodules based on multi-scale convolution neural network.A multi-scale feature fusion strategy is introduced on the deep convolutional network(VGG16)for large-scale image recognition.Through the fusion analysis of features on different scales,the accurate extraction of pulmonary nodules features is realized,and the semantic and detail information of pulmonary nodules features are retained greatly.The multi-scale fusion feature maps are used to detect pulmonary nodules,and the candidate boxes of pulmonary nodules at various scales are obtained.The non-maximum inhibition strategy is introduced to optimize the candidate boxes of pulmonary nodules at various scales,and the optimal location of pulmonary nodules detection is obtained.In this paper,the algorithm is verified by the datas of lung nodules from the Lung Imaging Database Consortium(LIDC-IDRI),and the average detection accuracy of lung nodules is 90.9%.2.Automatic detection algorithm of micro pulmonary nodules based on improved dense connection network.Aiming at the problems of difficult detection and incomplete feature extraction of micro pulmonary nodules,this paper constructs an automatic detection algorithm of micro pulmonary nodules based on improved dense connection network.By making full use of the advantages of convolutional dense connection in feature transfer and feature extraction,a feature extraction network with five dense connection modules is established in this paper to fully extract the multi-scale features of micro-pulmonary nodules and perform fusion analysis on the multi-scale features to improve the detection accuracy of micro-pulmonary nodules.The average detection accuracy of the improved dense connection network for micro pulmonary nodules is 91.64%,which is 1.16% higher than that of the VGG16 network with multi-scale information fusion.