Research On Multimodal Image Classification Based On Machine Learning

Multimodal magnetic resonance imaging can clearly show the lesions inside the human body.By analyzing MRI images,doctors can accurately grasp the patient’s condition,so as to formulate an appropriate treatment plan.Therefore,magnetic resonance imaging technology is widely used in the diagnosis process of various diseases.Glioma is a typical malignant tumor of the brain,which seriously threatens human life and health.Magnetic resonance imaging can be used to classify gliomas,but in the actual clinical data collection process,there are often various reasons that cause part or even the entire modalities of the multimodal data to be missing,which is not conducive to the doctor’s diagnosis.However,when the model is trained with missing multimodal data,the prediction effect of the model may be unstable due to the lack of key information.Endometrial carcinoma is a common gynecological malignant tumor.The traditional diagnosis method of the disease mainly relies on doctors to manually analyze magnetic resonance images,which is not only time-consuming and labor-intensive,but also it is difficult to directly know the depth of tumor invasion only by image analysis.In addition,there are few related researches on combining artificial intelligence with clinical data of endometrial cancer,and there is no mature method.In view of the above background and problems,this thesis mainly studies the completion and classification of brain glioma image missing data based on the tensor network model and the imaging staging of endometrial cancer.The main work and contributions include the following points:1.For clinical imaging data,systematically introduce the feature extraction,feature selection and classification methods used in this thesis,compare various data completion methods,and summarize and analyze the advantages of tensor completion methods based on low-rank hypothesis compared with traditional methods.2.A missing modality brain image classification framework based on tensor network model is proposed.The framework is based on the low-rank factor model of the tensor ring,decomposes the multimodal glioma feature data through the tensor ring to obtain a series of kernel tensors,and mines the hidden information between the features of different modalities.And by imposing low-rank constraints on the kernel tensor,the calculation amount is greatly reduced;finally,the completed data is put into the SVM classifier to realize the classification function.;finally,the completed data is put into the SVM classifier to realize the classification function.Compared with classification frameworks based on other completion methods,the framework proposed in this thesis has better performance.3.A machine learning-based framework for staging of endometrial cancer is proposed.The framework extracts first-order statistics,second-order texture and three-dimensional shape features from multimodal tumor images of patients,and uses various methods based on LASSO for feature selection.Finally,the selected features are put into different classifiers for classification.By comparing the results of different feature selection and classification methods,the combination with the best staging effect was obtained.Experiments have shown that the framework can basically achieve staging of endometrial cancer.