A deep learning radiomics model for preoperative grading in meningioma

Area of Science:

• Neuroimaging
• Machine Learning
• Oncology

Background:

• Meningiomas are common primary brain tumors.
• Accurate grading of meningiomas is crucial for treatment planning.
• Noninvasive methods for differentiating meningioma grades are needed.

Purpose of the Study:

• To develop and validate a deep learning radiomics (DLR) model for noninvasive meningioma grading.
• To assess the performance of the DLR model using routine post-contrast MRI data.
• To compare the DLR model's efficacy against traditional radiomic approaches.

Main Methods:

• 181 patients with histopathologically diagnosed meningiomas were included.
• Deep learning features were extracted from post-contrast T1-weighted MRI using a convolutional neural network.
• A DLR model was built using random forest feature selection and linear discriminant analysis classification.

Main Results:

• The DLR model, comprising 39 features, demonstrated strong discrimination in both primary and validation cohorts.
• In the validation cohort, the DLR model achieved an AUC of 0.811, with 0.769 sensitivity and 0.898 specificity.
• The DLR model outperformed a radiomic model based on hand-crafted features.

Conclusions:

• A DLR model utilizing routine MRI data can effectively predict meningioma grades noninvasively.
• The DLR model offers superior quantization capabilities compared to hand-crafted features.
• This model has the potential to inform clinical decisions regarding meningioma observation or treatment.