Application Of Deep Learning And Radiomics In The Diagnosis And Treatment Of Glioma

Glioma is a common type of brain tumor,which can be divided into low grade glioma(LGG)and high grade glioma(HGG).Magnetic resonance imaging(MRI)is often used for the clinical diagnosis of brain tumors due to its high soft tissue resolution and spatial resolution without radiation.Early interventional treatment can effectively improve the survival rate of patients.The treatment of glioma is mainly based on surgery.Segmentation of the tumor area is an important part of the surgical planning.The prognosis of different grades of glioma is different.However,the segmentation of tumors mainly relies on manual segmentation by doctors or experts,which is inefficient.Pathological results are the gold standard for determining tumor grade,but it is an invasive test that may cause secondary harm to the patient.In view of the above problems,this paper conducts research on the application of deep learning and radiomics in the auxiliary diagnosis and treatment of glioma.The main work content is as follows:(1)An MRI glioma segmentation algorithm based on dense connection blocks and Transformer,ADT-UNet,is proposed.Based on the U-Net network,the algorithm introduces dense connection structure and attention mechanism to enhance the feature extraction ability of the coding block.At the same time,the Transformer structure is added at the end of the encoder to further enrich the feature representation.In addition,the multiscale feature fusion module helps to capture fine-grained semantics and coarse-grained semantics at the full scale.Compared with the latest medical image segmentation algorithms,the proposed method obtained competitive results in each evaluation index.Ablation results show that the three modules can improve the segmentation accuracy of the model,and the combination of the three modules has the best segmentation efficiency.On the test set,the DSC,IOU,PPV and Sen were 0.972,0.945,0.970 and 0.973,respectively.(2)A prediction model for glioma grading based on MRI multi-sequence imaging is proposed.This method firstly extracted high-throughput radiological features of four kinds of conventional MRI sequence image data,then Spearman correlation coefficient method and LASSO regression algorithm were used for multi-step feature selection,and the top 10 best feature subsets were selected,and the four best feature subsets were fused.Eleven feature fusion subsets were generated.Finally,seven machine learning models were used for training.The experimental results show that the hierarchical prediction has the best performance when all the four sequence features are fused and the logistic regression model is used for modeling.In the test set,the model also showed good predictive performance,with AUC value,accuracy,sensitivity and specificity of 0.981,0.929,0.968 and 0.800,respectively.(3)A brain glioma auxiliary diagnosis and treatment system based on deep learning and radiomics was designed and implemented.The system could automatically segment tumor lesions and predict tumor grading,providing reliable diagnostic basis and effective lesion information for doctors to assist in the subsequent diagnosis and treatment process.