| Purpose:To explore the value of the radiomics model based on conventional multi-sequence MRI in predicting the pathological immunohistochemical Ki67,PR expression and histological grading of meningiomas before surgery,which is helpful for the formulation of clinical treatment plans and prediction of prognosis.Materials and methods:We retrospectively analyzed the MRI images（including four sequence images of T2 WI,T1WI,T2 WIFLAIR,and T1C）of 293 patients with meningioma.All patients were confirmed by pathology after surgery.Among them,there were 224 patients with Ki67<4%,69 patients with ≥4%,226 PR positive patients,67 PR negative patients,264 patients with lowgrade meningioma,and 29 patients with high-grade meningioma.36 cases of fibrous type and 257 cases of non-fibrous type.At the same time,some clinical and imaging features of meningioma patients were collected for analysis.Two radiologists with different seniority used ITK-SNAP software to delineate and segment the enhanced area of tumor lesions layer by layer.The resulting EnH ROI was then inflated by 3 mm on the uAI Research Portal radiomics platform to obtain the EnH 3mm ROI.Image feature extraction was performed on the ROI of each MRI sequence,and then the obtained radiomic features were screened and reduced in dimension.According to different tasks,the statistically significant radiomic features obtained by screening were used to build models by logistic regression,and the statistically significant radiomic features and clinical-imaging features were combined to establish a joint prediction model of Ki67 and PR expression.The evaluation applied receiver operating characteristic（ROC）curve and its related indicators to evaluate the performance of different prediction models.Results:A radiomics model for predicting Ki67,PR expression and histological grade of meningiomas using ROC assessment.In the Ki67 expression radiomics prediction model,the AUC of EnH model was 0.732,the sensitivity was 0.730,the specificity was 0.636,and the accuracy was0.708;the AUC of the EnH 3mm model was 0.767,the sensitivity was0.703,the specificity was 0.636,and the accuracy was 0.688;In the PR expression radiomics prediction model,the AUC of the EnH model was0.663,the sensitivity was 0.636,the specificity was 0.649,and the accuracy was 0.646;the AUC of the EnH 3mm model was 0.796,the sensitivity was0.818,the specificity was 0.757,and the accuracy was 0.771;In histological grading,the AUC of the EnH model was 0.902,the sensitivity was 0.830,the specificity was 0.828,and the accuracy was 0.829;the AUC of the EnH 3mm model was 0.928,the sensitivity was 0.864,the specificity was 0.828,and the accuracy was 0.860;the AUC of the EnH model in histological typing was 0.805,the sensitivity was 0.712,the specificity was0.694,and the accuracy was 0.710.was 0.893,the sensitivity was 0.868,the specificity was 0.806,and the accuracy was 0.860.Univariate analysis of clinical and imaging features found that tumor shape and enhancement degree were statistically significant for Ki67 expression,and tumor shape was statistically significant for PR expression.The clinical,imaging and radiomic features were screened to establish a joint prediction model.For the analysis of the joint prediction model,the AUC of the Ki67 expression joint prediction model was 0.779,the sensitivity was 0.703,the specificity was 0.636,and the accuracy was 0.688 The AUC of the PR expression joint prediction model was 0.799,the sensitivity was 0.818,the specificity was0.649,and the accuracy was 0.688.Conclusion:1.The radiomics model based on conventional multi-sequence MRI images has certain value in predicting the expression of Ki67,PR and histological grading of meningiomas before surgery.2.In clinical and imaging features,tumor shape and enhancement degree were significantly correlated with Ki67 expression,and tumor shape was also significantly correlated with PR expression.3.The performance of the radiomics prediction model based on EnH 3mm ROI is obviously better than that based on EnH ROI.4.The performance of the joint prediction model is better than that of the radiomics prediction model. |
