Research On CT Image Lung Nodule Detection Method Based On Deep Learning

With the increase in the number of smokers in the society and the deterioration of air quality,lung cancer has become one of the cancers that endanger human health,and the survival rate of lung cancer patients is extremely low.Lung cancer is usually characterized by pulmonary nodules,so if you can detect pulmona ry nodules and their cancer in time,early detection and early treatment will be of great value in improving the five-year survival rate of early lung cancer patients.Currently,physicians image the lung area using CT techniques to detect lung nodule lesions.However,doctors make qualitative judgments based on their own experience.Under the conditions of manual visual inspection,there are limitations such as low accuracy,poor repeatability,slow speed,high labor intensity,and subjective factors.Computer-aided diagnosis technology assists doctors in discovering lung nodules and making diagnosis through image processing technology,which reduces the workload of doctors.Therefore,it is necessary to design intelligent,standardized and digital CT image lung nodule defection methods,which helps doctors to screen lung nodules more quickly and effectively,shorten the patient’s diagnosis cycle,and help medical institutions to establish lung CT.Image database and lung cancer analysis and statistics to improve the treatment of lung nodules and lung cancer,optimize medical resource allocation.The difficulties in the detection of pulmonary nodules are:1.Interference of other tissues in the CT image of the lung;2.The cross section of the blood vessels in the lung parenchyma is similar to the cross section of the lung nodules,which may easily lead to false detection;3.The size of the lung nodules and the shape of the edges different from each other,it is difficult to extract features.Aiming at the problem of interference from other organs in lung CT images,this paper proposes a lung parenchyma segmentation algorithm based on the fusion of VGG-16 and dilated convolution.The dilated convolution can expand the receptive field without increasing the convolution parameters.The partial convolution layer in VGG-16 is replaced by the dilated convolution method,and the pooling layer is eliminated.At the same time,the hyper-column feature is used in the network for feature fusion.Finally,the pixels are classified by MLP to obtain the segmentation result.In this paper,the lung parenchyma region in the CT image is segmented to form a mask,which excludes the interference of other organs in the CT image on lung nodule detection.The experimental results show that the DSC coefficient obtained by the network is 0.9867,which is better than the image processing algorithm and other convolutional neural network algorithms.In view of the easy misdetection of lung nodules and the difficulty of feature extraction,this paper proposes a lung nodule detection algorithm based on the combination of Reception and Faster R-CNN.The ResNet can effectively alleviate the gradient due to the increase of network layers.The disappearance problem,the Inception structure can connect different convolutional layers in parallel,improving the utilization of internal resources.This paper combines the ResNet with the Inception structure,and proposes the Reception module to extract features,and then uses the Faster R-CNN structure for classification and regression.The experimental results show that the accuracy rate of the network in this paper reaches 92.8%and the expected effect is achieved.