Research On Lung Nodule Classification In CT Image Based On Self-supervised Learning And Transfer Learning

Lung cancer is currently one of the most threatening cancers to human health.In order to reduce the burden on doctors,and improving the diagnosis efficiency of lung cancer treatment,deep learning has been introduced into the clinic to judge benign and malignant lung nodules.Deep learning learns the inherent characteristics by fitting a large amount of data.At the same time,the larger the data scale,the better the performance of the model.Due to the particularity of the medical field,it is usually difficult to obtain large-scale labeled data sets.Self-supervised learning does not require data to have labels,and relying on the characteristics of the data to train the model will help improve the performance of the model.Aiming at the problem of categorizing benign and malignant lung nodules in CT images,the current self-supervised learning method only uses unlabeled lung CT data to train the model,ignoring the use with lung CT data in the selfsupervised learning process.The existing methods also ignore how to use CT data sets of other tissues and organs other than the lungs to realize self-supervised learning.To solve the problems mentioned above,this thesis does some intensive research on the classification of lung nodules based on self-supervised learning and transfer learning.The main work is as follows:(1)Aiming at the problem of self-supervised learning using only the unannotated CT image data set of the lungs,this thesis designs a self-supervised learning model that combines part of the annotation information.With a single-input and dual-output structure,allowing CT images containing lung nodules and CT images without lung nodules to share the feature extraction network,and the models were obtained training results from the two output ends.The weighted loss function is used to adjust the weights of the two types of data,and finally the parameter migration of the feature extraction network is used to classify lung nodules.Experiments on the LIDC data set show that training the model by adding data containing lung nodules enables the network to learn more important features related to the target task while meeting the large amount of data required for training.(2)Regarding how to use CT image data sets of various tissues and organs for selfsupervised learning,and pre-train a model with strong generalization ability to improve the classification performance of lung nodules.This thesis designs a self-supervised learning framework based on multi-organ CT images.Since there are certain differences between different organs,if a single convolution kernel is used directly for feature extraction,the extracted features will not be comprehensive enough.Therefore,for collecting local features,as well as global ones,a multi-scale convolution kernel is employed,and then the channel attention mechanism is used to select feature channels with higher importance.With the aim of narrowing the impact from the gap between the source domain and target domain data,this thesis joins the domain adaptive network in the process of self-supervised learning model migration,so as to effectively improve the accuracy of lung nodule classification through self-supervised learning.