Detection And Segmentation Of Lung Tumors Based On Deep Neural Network

Lung cancer has become one of the diseases with the highest morbidity and mortality worldwide.Low-dose CT images play a very important role in the treatment of the disease,and doctors can use the images to reflect the feature information to formulate a treatment plan.However,because reading CT images can take a lot of time and energy,and because doctors may be inexperienced and fatigued,it is important to understand the characteristics of CT images.This leads to missed or misdiagnosis.In addition,as further treatment usually requires more detailed information about the tumor,manual segmentation is time-consuming.To address this problem,computer-aided diagnosis can effectively save doctors’ time in reading films and improve efficiency and accuracy.A set of deep neural network based detection was designed based on the CT image data to study the various characteristics of tumors.with segmentation algorithm,which finally gives information about tumor location and size to assist doctors in treatment planning.In this thesis,a novel three-dimensional convolutional neural network detection algorithm with residual structure is designed.The network uses three-dimensional convolution to adequately extract the three-dimensional information from the image,and the residual structure reduces the information loss between layers.The proportion of actual labeled boxes is calculated by the K-medoids method,which is useful for the training of regional recommendation networks to get more Accurate location information.A group normalization method and Momentum’s gradient descent strategy method were additionally added to shorten the convergence time of the training network.A tumor recall of 0.905 was achieved on the test set.A three-dimensional convolutional full neural network with residual structure was designed for the tumor segmentation requirement,which mainly integrates the global information with the Local information,low and high level features,is achieved in the segmentation of the detection network given to the solid and non-solid tumors Dice score of 0.82.The segmentation of the tumor can be better accomplished to achieve the goal of assisting doctors in formulating further treatment plans.In summary,the main work of this thesis is to investigate the deep neural network-based algorithm for the detection and segmentation of lung tumors,which achieves the detection and segmentation of tumors from the entire process from image input to information output requires no human intervention,which can effectively reduce the burden on doctors and improve their efficiency.