Research On Lung CT Image Based On Deep Learning And Radiomics

Due to people’s irregular lives,unhealthy diet and smoking,lung cancer has long been the cancer with the highest mortality rate.With the advancement of imaging technology,medical image analysis is now playing a very critical role in the process of early disease screening and diagnosis.Currently,the data of medical images continues to increase,and so does the pressure on radiologists,often leads to missed diagnosis and misdiagnosis.Over the last few decades,artificial intelligence has been making more and more contributions to the medical field,boosting new breakthroughs in intelligent analysis of medical images.Radiomics is used to extract deep information from a large number of medical images,obtain a huge number of features,and then build a machine learning model to assist therapists in making diagnoses and improve work efficiency.Medical image segmentation is a key part of radiomics,and the accuracy of nidus segmentation largely influences the diagnosis.This thesis tries to study the segmentation of pulmonary nodules on lung CT images.In order to effectively prevent the spread and metastasis of lung cancer cells,this thesis also tries to construct a model to predict whether patients with lung cancer will develop distant metastasis,so as to help therapists make analysis and treatment plans.The main research work is as follows:(1)Since the U-Net and U-Net++ algorithms fail to effectively use the feature weight information,the segmentation boundary appears under-segmented in the complex scenes of medical images.To solve this problem,this thesis puts forward an improved UNet++ network model based on the fusion adaptive weighted aggregation strategy,and applies it to pulmonary nodule segmentation on CT images.The model extracts semantic information at different depth features in the convolutional neural network,combines the weight aggregation module to automatically learn the weights of the features of each layer,and then loads the learned weights onto the segmentation maps obtained by sampling on each feature layer to obtain the final segmentation result.This algorithm effectively improves image segmentation performance.The method in this thesis can achieve accurate segmentation on the tiny details of the tumor,and better solve the problem of under-segmentation when the pulmonary nodules grow infiltratively to the surrounding area.(2)Patients with non-small cell lung cancer are more likely to develop distant metastasis,and the metastasis of cancer cells has a serious impact on the patients’ physical functions.Hence,the ability to make accurate distant metastasis prediction is vital in subsequent clinical treatment.Through analyzing 356 pieces of data collected from such patients in the Chongqing University Cancer Hospital,this thesis standardizes the image processing,extracts radiomics features from nidus on CT images.The clinical-semantic features are also extracted from the patient’s clinical data and diagnostic information,then the feature selection is performed,and the radiomics model,clinical-semantic model and joint model are respectively constructed through logistic regression algorithm to predict distant metastasis.Besides,this thesis draws the nomogram of the combined model to evaluate the risk of distant metastasis of patients and help therapists make better clinical decisions.