Research On Auxiliary Diagnosis Algorithm Of Pulmonary Nodules Based On Self-supervised Model Migration

At present,the number of deaths and new incidence of lung cancer in China temporarily ranks the first place.Observing lung nodules in patients’ lung areas through low-dose CT is recognized as the most effective means of early screening for lung cancer,which can effectively reduce the mortality of patients with lung cancer.Traditional methods of early screening for lung cancer are time-consuming and laborious,and easy to miss and misdetect.However,the computer aided diagnostic system of lung nodules based on deep learning and machine learning methods can realize rapid detection and diagnosis of lung nodules.In the training of the pulmonary nodular aided diagnostic model,there are still some problems,such as unbalanced distribution of the size of pulmonary nodules,small number of labeled data sets and unbalanced number of positive and negative samples.Aiming at the segmentation and false positive reduction classification tasks of pulmonary nodules in the aided diagnostic algorithm of pulmonary nodules,this thesis proposes a solution based on self-supervised model migration,and the following research work was carried out:1.Self-supervised model framework and agent task design.This thesis combined the advantages of generative and discriminant self-supervised learning to build a selfsupervised model framework,and based on unlabeled lung CT image training.The autoencoder structure model can learn the basic features of lung,and five kinds of proxy tasks based on image transformation can learn the specific features of lung from multiple angles.The self-supervised model can be transferred to the task of pulmonary nodule segmentation in Chapter 4 and false positive classification in Chapter 5 as a pre-training model to improve the performance of subsequent target tasks.The experimental results show that the self-supervised model can effectively improve the training convergence speed and performance of subsequent target tasks.2.Segmentation algorithm of pulmonary nodules based on self-supervised model migration.In the thesis,the 3D-UNet model is improved based on multiple residual modules and void convolution skip connections.Meanwhile,the loss function is replaced by the Tversky loss function of = 0.2 in the training stage.The improved model can extract the feature information of pulmonary nodules from different scales without increasing the number of excessive parameters,and enlarge the receptive field of the low-level feature map.A nodule segmentation model trained with = 0.2 Tversky loss function showed high sensitivity without losing too much segmentation accuracy.Experimental results show that the segmentation model combined with various strategies in this thesis achieves 73.56% segmentation accuracy and 82.73% sensitivity,which can effectively segment pulmonary nodules of different sizes and shapes and reduce missed detection of pulmonary nodules.3.Research on false positive classification algorithm of pulmonary nodules based on self-supervised model migration.In view of the extremely unbalanced positive and negative samples of false positive classification tasks,CT-GAN network was used in this thesis to expand positive samples and sample negative samples,so as to make the distribution of training data more balanced.Aiming at the problems of different scales and shapes of pulmonary nodules,in the thesis,the false positive reduction classification model based on progressive feature extraction module and self-supervised model encoder can adaptively adjust the level and depth of feature abstraction when dealing with pulmonary nodules of different sizes,and effectively distinguish different categories of nodules.Experimental results show that the model in this thesis achieves an average Competitive Performance Metric value of 0.882 on the LUNA16 dataset,which can effectively realize the false positive screening of pulmonary nodules and improve the auxiliary diagnosis efficiency of pulmonary nodules.