Brain Invasive MRI Radiomics Of Meningioma With Peritumoral Edema

Objective: To investigate the value of MRI radiomics in predicting brain invasion of meningioma with peritumoral edema.Methods: We retrospectively collected 73 patients with pathologically confirmed meningioma and peritumoral edema on preoperative T2WI（20 patients with brain invasion and 53 patients without brain invasion）who underwent preoperative brain MRI from April 2017 to June 2021 in the Department of Radiology,the First Hospital of Jilin University and randomly divided into cross-validation group（n = 51）and test group（n = 22）according to the ratio of 7:3.We used ITK-SNAP to manually segment the tumor on T1 WI + C images and peritumoral edema on T2 WI images.Edema volume and tumor volume were extracted by RIAS as clinical parameters.On the basis of T1WI+C tumour segmentation,the brain-tumour interface was then automatically segmented using RIAS software to form an 8 mm wide brain-tumour interface,and the histological features of the T1WI+C tumour,the braintumour interface and the used images of both were extracted separately,and the histological features closely associated with brain invasion were selected using LASSO.Then LR models of the tumour,brain-tumour interface and the imaging histological features of the two fused were constructed separately,and the optimal imaging histological model combined with statistically significant clinical parameters was selected to form a combined clinical-imaging histological GBDT model.AUC values,sensitivity,specificity,precision,recall,F1 score and accuracy were calculated using ROC curves to assess the efficacy of the models.Then LR models of the tumour,brain-tumour interface and the imaging histological features of the two fused were constructed separately,and the optimal imaging histological model combined with statistically significant clinical parameters was selected to form a combined clinical-imaging histological GBDT model.AUC values,sensitivity,specificity,precision,recall,F1 score and accuracy were calculated using ROC curves to assess the efficacy of the models.Results: There were significant differences between meningiomas with and without brain invasion in terms of age,tumor location,peritumor edema volume,edema index,and WHO classification（P<0.05）,but not in terms of gender,tumor volume,or Ki-67 expression level（P>0.05）,and a clinical LR model was developed using general clinical parameters that were statistically different preoperatively（Cross-validation AUC = 0.916;Test AUC = 0.722）.Among the three imaging LR models,the T1 WI + C brain tumor and brain-tumor interface imaging fusion model had the best performance（Cross-validation AUC = 0.913;Test AUC = 0.861）,which significantly improved the predictive power compared with the brain tumor and brain-tumor interface imaging models alone.The clinical-imaging combined GBDT model was established after combining the clinical LR model on the basis of the imaging fusion LR model,and the AUC value of this model was improved（Cross-validation AUC = 1.00;Test AUC =0.889）,but the specificity and accuracy of the model were slightly decreased.Conclusion: The radiomics fusion model based on meningioma and brain-tumor interface is superior to the meningioma radiomics model alone and the brain-tumor interface imaging model,coupled with the combined clinical-radiomics model constructed by clinical features to further improve the predictive ability of the model.