Research On Pulmonary Nodules Segmentation Based On Multi-scale Densely Connection 3DU-Net

The incidence rate and mortality rate of lung cancer are the highest in the world.Lung cancer is a serious threat to human health.Early detection and treatment are important to reduce lung cancer mortality.Pulmonary nodules are the early forms of lung cancer,the diagnosis of pulmonary nodules is the key to improve the cure rate of patients with lung cancer.CT is the most widely used diagnosis and treatment,however the number of CT data is growing rapidly.Therefore,it requires a huge work for doctors to observe CT images for disease diagnosis,and the efficiency is low,the diagnosis results are usually subjective.In order to improve the diagnosis efficiency and reduce the work of doctors,it is important to segment pulmonary nodules in medical images by artificial intelligence.How to segment CT image efficiently and accurately based on artificial intelligence is the focus of this thesis.U-Net is a typical deep neural network,because of its simple structure and strong generalization ability,it has become a common network in medical image segmentation.With the deepening of network depth,the gradient of U-Net will disappear,and the feature between different scales is not fully utilized.To solve these problems,this thesis proposes a pulmonary nodules segmentation based on Multi-scale Densely Connection 3DU-Net(MSDU-Net).Based on U-Net,this thesis proposes a multi-scale densely connection module,which can fuse the features of different scales in the network together as the input of this stage,so that the input information obtained in each stage of the network is all the features of the previous stage,so as to improve the segmentation accuracy of the network.In addition,the residual 3D convolution(R3D-conv)module is introduced to combine U-Net with residual network,which can solve the gradient vanishing problem caused by the deepening of layers in deep convolution network.In this thesis,we embed R3D-conv module and MSD module in the U-Net to build a new network model,namely MSDU-Net.In order to verify the segmentation accuracy and generalization ability of our proposed MSDU-Net,we trained and tested it on two classical medical image datasets,Montgomery County X-ray dataset and Luna16 dataset,and compared it with some other excellent medical image segmentation methods.The experimental results show that the proposed MSDU-Net can achieve better segmentation results and its performance is generally better than other comparison methods.