Research And Implementation Of Lung Nodule Detection And Classification Method Based On Convolutional Neural Network

Nowadays,more and more people suffer from lung cancer,which is one of the serious threats to human survival in the world.Lung nodules is the main markers of early lung cancer.Medical experts believe that early diagnosis of suspected lung nodules by computed tomography scans can give patients with lung cancer the hope to survive.With the development of CT technology,the explosion of CT data makes doctors’ work more and more heavy.So computer-aided diagnosis technology has become a powerful tool for doctors in the diagnosis process.However,some characteristics of pulmonary nodules,such as their size and surrounding environment,make existing pulmonary nodular detection and diagnosis systems not perfect as expected.However,recent studies have shown that convolutional neural network has a powerful ability in the in the field of image processing promoting the development of medical image processing.In short,in order to improve the model’s detection and diagnosis capabilities and make it practical as soon as possible,We propose a method to improve the original lung nodule detection and diagnosis system.The following is the main research part:(1)Lung nodule detection based on Faster-RCNNTo take advantage of the hierarchical nature of the network,the maximum pooling operation is used to keep the size of the first and middle layers in VGG16 to keep the same size as the last layer feature map,Then the features of Conv1,Conv3,and Conv5 form a new feature.Next,512 1*1 convolution kernels are used,so that the dimension of the newly combined feature map becomes 512.The K-means algorithm is used to select and modify the size and proportion of anchor points in the RPN network.To speed up the training of the model,this article replaces the 3*3 convolution in the original RPN network is replaced with a depth separable convolution.The test on the LIDC-IDRI data set shows that the average accuracy is 0.91,which is higher than the same type of algorithms compared in the experiment,providing a new research idea for lung nodule detection algorithms.(2)Classification algorithm based on VGG16In order to improve the classification ability of the benign and malignant pulmonary nodule classification model,a nodule classification algorithm is proposed based on VGG16.First,this article integrates the doctor’s diagnosis process and extracts the 50-dimensional prior features of lung nodules.With the view of noticing the problems surrounding the lung nodules during the diagnosis process,this paper designs a network structure,which is named multiple-input single-output.At the same time,this paper replaces the pooling layer in the original VGG16 with a convolutional layer.Being aimed at reducing the complexity of the model,this paper adds a dual-path connection block to VGG16.Finally,this paper fuses the extracted 50-dimensional prior features with the features proposed by the model.The ROC,based on Luna16 dataset,is 0.92,which shows that the idea(multiple-input single-output mode l)makes sense.(3)Designing a lung nodule detection and benign and malignant classification system based on the TKINTER framework,which can help doctors reduce their workload.