Research On Detection Of Pulmonary Nodules Based On CT Image Sequence

In recent years,lung cancer has been a major problem plaguing our country and even the world.Its morbidity and mortality are high,causing thousands of deaths each year.At present,the most effective method for lung cancer is still early detection and early treatment.The most obvious sign of early lung cancer is pulmonary nodules,and computed tomography(CT)is the most important means of detection.However,CT sequence image data are numerous,and with the increase of patients in recent years,CT image data is increasing day by day,which leads to the increasing burden of radiologists.Because the pulmonary nodules are small and difficult to observe,and the structure of the lung is complicated,it is difficult for radiologists to ensure that they can read films with high intensity for a long time without making mistakes.This urgently requires the calculation of auxiliary diagnosis to reduce the burden on the physician and increase the accuracy of the detection.With the rapid development of deep learning in recent years,the algorithms in the field of image processing have ushered in tremendous changes,and the accuracy of algorithms in many sub-domains has been further improved.In this thesis,the deep learning algorithm is used to study the algorithm of lung parenchymal segmentation and lung nodule detection.The advantages and disadvantages of the current algorithm are compared and analyzed,and the corresponding improvements are proposed in the algorithm,which increases the accuracy of the algorithm and improves the speed of the algorithm.In this thesis,we propose a method to find regions of interest for separating left and right lungs,which speeds up the algorithm.In the stage of feature extraction,a multi-scale feature of the pulmonary nodules based on the feature pyramid of the dilation convolution is proposed to improve the accuracy of the model.In the region proposal stage,the size of candidate box is improved to meet the size of pulmonary nodules in this thesis.In view of the imbalance of positive and negative samples,the loss function is improved to improve the detection accuracy.In addition,three-dimensional convolution neural network is used to classify the final detection results and eliminate false positives.Meanwhile,region proposal network is used to generate three-dimensional image blocks to enhance the generalization ability of classification network.In this thesis,based on the actual CT data of pulmonary nodules,the detection results of the algorithm are compared and verified,which proves the effectiveness of the algorithm.This thesis proposes a method to complete the whole detection process with a smaller network.Compared with the large network,this method has faster application speed and saves hardware cost.It has important engineering application value.