Cancer Imaging Analysis Based On Deep Learning And Radiomics

With the rapid development of medical imaging technology,medical imaging plays an important role in clinical diagnosis and treatment.How to quickly and effectively mine information from medical images and make correct clinical decisions is the key to improving the cure and survival rate of patients.Medical imaging data is increasing in a large amount while the number of radiologists is growing relatively slowly.Doctors face huge workload and pressure,which are prone to misdiagnosis and misdiagnosis.In recent years,the development of artificial intelligence technology has provided new ideas for solving such shortages of medical resources.Especially radiomics,which integrates the knowledge of disciplines such as radiology,oncology,and computer science.It can detect high-throughput features from medical images and perform statistical modeling to make clinical decisions.The pipeline of radiomics usually includes image data acquisition,tumor segmentation,feature extraction,feature selection,and modeling analysis.Among them,tumor segmentation is a critical step in radiomics research.Due to the variety of medical images and blurred tumor boundaries,accurate segmentation is a great challenge.Manual segmentation by doctors is time-consuming and laborious,and suffers subjectivity,resulting in poor reproducibility of radiomics research.Besides,many radiomics studies only focus on small samples,resulting in poor generalization performance.The above problems,to a great extent,limit the application of radiomics in clinical practiceIn order to solve the problems,this thesis focuses on the study of radiomics for brain tumor and lung cancer based on magnetic resonance imaging(MRI)and computer tomography(CT).We comprehensively apply the deep convolutional neural network(CNN)and traditional machine learning technology to solve the key problems in radiomics.The main contributions and innovations of this thesis are as follows1)Design of neural network for MRI brain tumor segmentation.Brain tumors vary significantly in shape and size,and their boundaries are relatively blurry.Even manual segmentation by doctors is challenging.To solve this problem,we propose a trilateral segmentation network(TriSegNet)based on deep learning.TriSegNet consists of three paths,including a spatial path to capture rich spatial information,a context path to obtain a sufficiently large receptive field and encode rich context information and a localization path to restore the original input size and restore the tumor details.This network can learn high-discriminatory semantic features while maintaining a high spatial resolution to effectively identify smaller tumor regions and blurry oundaries.Experimental results show that brain tumors can be segmented accurately and effectively using the proposed method.2)Design of neural network for CT small cell lung cancer segmentation.CT images often have higher resolution than MRI images,and how to effectively segment CT images is very challenging.To solve this problem,we propose a hybrid segmentation network(HSN)based on deep learning.HSN contains a lightweight 3D CNN and a fine-grained 2D CNN.3D CNN uses down-sampled images and separable spatial/temporal 3D convolutions to reduce memory requirement and calculation cost.2D CNN can maintain a higher spatial resolution to learning fine-grained semantic information.Then we propose a hybrid feature fusion module to effectively fuse 2D and 3D features.This network combines the advantages of 3D CNN learning long-range context information and 2D CNN learning fine-grained semantic information.Experimental results show that cancer regions can be segmented accurately and effectively using the proposed method.3)Automated brain tumor grading based on radiomics analysis.This method combines MRI brain tumor automated segmentation and radiomics modeling analysis.First,the multi-scale 3D convolutional neural network is used to segment the entire tumor area,and a large number of radiomic features are extracted.The optimal feature subset is selected by recursive feature elimination method,and three classifiers are evaluated.The experimental results show that the radiomics analysis can effectively distinguish the high-level and low-level brain tumors and assist doctors in making clinical decisions to a certain extent.4)Prediction of the chemotherapy effect of small cell lung cancer based on radiomics analysis.High throughput radiomic features are extracted from the lung cancer area,and the 13 most discriminative features are selected.The support vector machine algorithm is used to obtain a radiomics score,which is combined with clinical features to construct a model.Experimental results show that the model can effectively predict the clinical response of small cell lung cancer patients to first-line chemotherapy.5)Non-small cell lung cancer subtype differentiation based on radiomics analysis.The histological phenotype identification of non-small cell lung cancer is crucial to the choice of the treatment plan.The number of samples used in many current studies is relatively small,and there is a lack of independent external verification dataset.Therefore,the generalization ability of the model cannot be fully verified.To solve these problems,we retrospectively study three independent datasets from different centers.For each dataset and a combination of multiple datasets,a corresponding radiomics model is constructed.The experimental results show that the model based on the merged dataset is more stable than the model obtained from a single dataset.