Machine learning based radiomics outperforms tumor/cerebellar peduncle T2-weighted imaging intensity ratios and ROIs-based apparent diffusion coefficient measurements in the preoperative prediction of meningioma consistency.

PURPOSE: Accurate preoperative assessment of tumor consistency is critical for surgical outcome in meningioma patients. This study aims to confirm whether machine learning (ML) based radiomics demonstrates superior performance compared to conventional methods, including tumor/cerebellar peduncle T2-weighted imaging intensity (TCTI) ratios and manual regions of interest-based apparent diffusion coefficient (ROIs-based ADC) measurements, in predicting meningioma consistency.

MATERIAL AND

METHODS

168 meningioma patients were enrolled in this study. Referring to the Zada's consistency grading system, meningioma consistency was classified into three categories: soft (grades 1 and 2), moderate (grade 3), and hard (grades 4 and 5). TCTI ratios were calculated and ADC values were measured. Radiomics features were extracted from post-contrast T1WI, T2WI, and ADC. Predicting models were constructed using support vector machine, and predictive performance was evaluated by the area under the receiver operating characteristic curve (AUC).

RESULTS

30 meningiomas were classified as soft, 92 moderate, and 46 hard. For "soft" prediction, the best AUC values of ROIs-based ADC measurements and TCTI ratios were 0.56 and 0.74 in the validation set, respectively. In contrast, radiomics model achieved an AUC of 0.88 in the validation set. For "hard" prediction, the best AUC values of ROIs-based ADC measurements and TCTI ratios were 0.63 and 0.69 respectively for the validation set. The radiomics model had an AUC value of 0.79 in the validation set.

CONCLUSION

Radiomics models outperform ROIs-based ADC measurements and TCTI ratios in the non-invasive prediction of meningioma consistency. Our results may promote the applications of ML techniques in clinical practices.