Predicting Intracranial Tumor Grade By Quantitative Radiomic Features

In recent years,the incidence of intracranial tumors which are the tumors of extra cerebral location has been on the rise,accounting for about 5%of total body tumors.High-grade tumors usually exhibit aggressive growth behaviors with a high rate of local recurrence,and there is a statistically significant difference in the prognosis of patients with high-grade and low-grade tumors.Therefore,accurate tumor grading is of great importance,which affects treatment planning and prognosis of patients.However,conventional imaging is not able to accurately distinguish between high-and low-grade tumors and the diagnosis of high-and low-grade tumors still depends on the histopathological biopsy from surgical resection.Histopathological diagnosis is an invasive method with many limitations,especially the inability to evaluate the heterogeneity of the entire tumor.Radiomics,a new technology,can extract high-throughout characteristics reflecting tumor heterogeneity from medical images,and achieve the preoperative diagnosis and prognosis prediction of tumor diseases by establishing the valuable radiomics signature.This non-invasive technique has great potential in assisting clinical decision-making and improving prognosis,thus facilitating counseling at an earlier stage of clinical care.In this paper,the theory and experiments of radiomics models were discussed in details,and the results of experimental verification and performance evaluation for MRI-based meningioma and glioma grading were shown respectively.In the experiment of meningioma grading,hand-crafted features and deep learning features were extracted and the deep learning features based radiomics model was constructed and compared with the hand-crafted features based radiomics model.The results showed the deep learning features had a stronger ability in characterizing meningioma grades.In the experiment of glioma grading,we built four radiomics models and got four corresponding radiomics signatures based on four MR sequences.Afterwards,a fusion model combining four radiomics signatures was constructed using the method of logistic regression.The results showed the fusion model had a stronger ability in grading meningioma than the model based on single MR sequence.In summary,radiomics model could achieve the accurate and pre-operative prediction of meningioma and glioma grading,which could provide quantitative basis and clinical assistance for clinical decision-making.