Lung Nodule Segmentation Algorithm In CT Images Based On Deep Learning

Lung cancer has always been a malady that mankind can not get over,Which threatens to human life and health.Pulmonary nodules,as the primary feature of primary lung cancer,are particularly important for the early detection and early diagnosis of lung cancer,which in turn can save the lives of countless patients.And computer-aided detection can help doctors to better test the patient’s condition,helpful to relieve the doctor’s busy work.Therefore,the use of CAD technology to segment the pulmonary nodules is of great significance.Compared with CT images,the pulmonary nodule are very small and very confusing with other tissues,there are still many problems in the study of pulmonary nodule segmentation algorithms.In this thesis,the existing methods are fully summarized and studied,and an efficient pulmonary nodule segmentation algorithm based on depth learning is proposed.Compared with machine learning,deep learning does not require manual feature extraction.Therefore,we mainly use the method based on U-NET to segment lung nodules,and the network in medicine Image processing problem is widely applied,compared with other networks has a simple structure,sample demand is relatively low and a series of advantages.Here we design the improved U-NET network,which mainly includes the design of dataset,the design of convolution layer,the design of pooling layer and the design of up-sampling layer.There is also the design of the residual network.To a certain extent,the residual network deepens the depth of the network and makes the training effect of the network better.Finally,the design of batch normalization can speed up the network training speed.Our specific processes mainly include image preprocessing,pulmonary parenchyma segmentation,the use of improved U-NET network model,Finally,the use of conditional random field to optimize the forecast results.Specifically,The main use of the lung segmentation is the morphological method to get the lung parenchyma,which is,our region of interest.Morphological methods are relatively simple and efficient,fast.Finally use our improved network for training and predicting.Experiments show that the method proposed in this thesis can be better segmentation of nodules,the algorithm improved accuracy.