Research On Method Of Pulmonary Nodules Detection Based On Convolutional Neural Network

Lung cancer is now known as the number one cancer worldwide and its incidence rate has increased year by year,especially in China.The initial symptoms of lung cancer are represented by pulmonary nodules.Clinically,detection of pulmonary nodules is the first step of lung cancer screening.At present,detection of pulmonary nodules mainly uses low-dose CT images to check.The pulmonary nodules in CT images are characterized by a variety of types,different structures,and small volumes.Even for experienced radiologists,it is a very time-consuming and laborious task to observe a large number of CT slice images and find all the nodules correctly by naked eyes,and it is easy to miss diagnosis or misdiagnosis.With the rapid development of deep learning in recent years,deep convolution neural networks have achieved great success in the field of computer vision,and it has also been more and more widely used in the field of medical images.At present,various types of medical image data have been accumulated to a certain extent,which brings a new development space for the use of deep convolution neural network to realize automatic analysis of medical images and assist doctors to achieve highprecision intelligent diagnosis of diseases.In this paper,RetinaNet is used to detect pulmonary nodules,and in order to improve its effect on the detection of lung nodules,a new method of detecting lung nodules based on improved RetinaNet is proposed.In order to make the set anchor parameters more suitable for pulmonary nodule detection task,K-means clustering analysis algorithm is used to cluster the size of the bounding box in the training set to obtain the appropriate anchor size;in order to make the network pay more attention to the relevant information of pulmonary nodule and ignore the irrelevant information through its own learning,so as to improve the ability of detecting pulmonary nodule,attention mechanism module is introduced;in order to make the features used for pulmonary nodule detection in the network have strong semantics and clear shallow feature information,and integrate more features at the same time,replace the feature pyramid in the original network with a bidirectional feature pyramid;finally,in order to reduce the amount of parameters and calculation of the improved retinanet,a lighter backbone network is introduced and the standard convolution of the rest of the network is replaced by depthwise separable convolution.The experimental results show that the improved RetinaNet based pulmonary nodule detection method proposed in this paper improves the detection effect of pulmonary nodule significantly,and after replacing the lighter backbone network and the depthwise separable convolution,it reduces the amount of model parameters and improves the running speed of the model when the detection effect changes little.