Pulmonary Nodule Detection Based On Improved CenterNet

As a cancer with high morbidity and mortality,early diagnosis and timely treatment are the most effective way to reduce the mortality rate of lung cancer in order to improve the survival chance of patients.Pulmonary nodules,as one of the important symptoms in the early stage of lung cancer,play a key role in the diagnosis and treatment of early stage lung cancer.Therefore,improving the detection performance of lung nodules is of vital clinical value.Target detection algorithm based on deep learning has been widely used in the detection of lung nodules.Due to the physical characteristics of lung nodules such as small volume,unpredictable position and complex lung texture,lung nodules detection is somewhat difficult.Therefore,there is still some room for improvement in the detection accuracy of lung nodules detection.This paper mainly carries out experimental research from the following aspects:(1)A CBAM-CenterNet lung nodule detection algorithm is proposed,aiming at the problems of the CenterNet algorithm with fewer layers in feature extraction network and easy loss of lung nodule feature information.ResNet50 is used as the feature extraction network of CenterNet algorithm.On the basis of this,CBAM attention module is added to build ResNet50+CBAM module,and then embedded into the CenterNet algorithm framework.In this way,the improved CBAM-CenterNet algorithm can promote the extraction of lung nodule feature information of different layers to a large extent,and at the same time,those insignificant lung nodule information can be ignored,so as to extract more abundant lung nodule feature information.In this paper,representative evaluation indexes such as mAP,Recall and Precision are used for experimental comparison.Under the same experimental conditions,the results show that the CBAM-CenterNet algorithm proposed in this paper has good detection performance of pulmonary nodules,which is significantly better than other comparison algorithms,with an average accuracy of 86.93%.Compared with the original CenterNet model,it is improved by 1.96%,which confirmed that the improved CBAM-CenterNet algorithm can effectively improve the detection effect of pulmonary nodules.(2)Due to the small amount of lung nodule data and the weak generalization and recognition ability of CenterNet algorithm,the data augmentation method is adopted to expand the lung nodule data set in order to improve the accuracy of the detection model.In this paper,four amplification methods including flip,contrast processing,Gaussian blur and color inversion were selected to expand the data set of the original 606 lung nodule images,and the data set after data augmentation had a total of 2 424 lung nodule images.To solve the problem that the improved ResNet50+CBAM feature extraction network is not highly accurate in the detection of small pulmonary nodules,the FPN CenterNet lung nodules detection algorithm is proposed.FPN is added to the ResNet50+CBAM feature extraction network to construct the ResNet50+CBAM+FPN module,so that the improved FPN-CenterNet algorithm can greatly improve the extraction of small-size lung nodule feature information.Under the same experimental conditions,the ablation results show that the FPN-CenterNet algorithm proposed in this paper has improved the detection effect,with an average accuracy of 89.90%,which is 4.93% higher than that of the original CenterNet model.It is confirmed that the improved FPN-CenterNet algorithm can effectively improve the detection effect of pulmonary nodules.In this paper,based on the above methods,CenterNet pulmonary nodule detection is experimentally verified,which confirms that the improved algorithm can effectively improve the detection effect of pulmonary nodules.