Radiomics-based shape and texture analyses on multiparametric MR imaging data for grading of biopsy-proven meningiomas

Purpose: Higher meningioma grades are associated with tumor growth and recurrence. Especially differentiation of low and intermediate grades is challenging, and reliable grading is not established. We applied radiomics-based shape and texture analyses on routine MRI for grading. Methods: MRI (T1/T2, T1 contrast-enhanced [T1CE], FLAIR, DWI, ADC) of n = 46 grade I and n = 25 II non-treated meningiomas with histological work-up were included. Manual segmentations were performed on FLAIR, T1CE and ADC by two radiologists in a consensus reading. Imaging data was preprocessed. Pyradiomics-package (vanGriethuysen2017) generated 815 radiomics features. Step-wise dimension reduction and feature selection were performed. Biopsy results were used as reference. Results: Four independent radiomics features where identified showing the strongest predictive values for higher tumor grades: roundnessFLAIR-shape (area-under-curve [AUC]: 0.80), cluster-shades-FLAIR/ T1CE (0.80), DWI/ADC-variability (0.72), FLAIR/T1CE-energy (0.76, p < 0.001 each). These features led in a multivariate regression model to an AUC of 0.91 for differentiation of grade I and II meningiomas. Conclusion: Radiomics applied on routine MRI is feasible for differentiation between low and intermediate meningiomas and a multivariate regression model yielded very good classification performance. In line with previous studies, higher meningioma grades were associated with shape parameters, contrast-enhancement and DWI/ADC-variability.