Interpretable MRI Radiomics for Preoperative Meningioma Consistency Prediction.

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Radiomic models for meningioma consistency prediction often lack interpretability, hindering clinical adoption.
• Key challenges include identifying predictive features, understanding spatial patterns, and correlating them with MRI characteristics.

Purpose of the Study:

• To develop an interpretable radiomic framework for predicting meningioma consistency using preoperative T1-Gd MRI.
• To enhance the clinical utility of radiomics by linking predictive features to spatial tumor patterns and MRI characteristics.

Main Methods:

• A cohort of 42 meningiomas was analyzed using preoperative T1-Gd MRI.
• An interpretable radiomic framework was developed, incorporating stable feature identification, SHAP attribution, and voxel-wise local radiomic mapping.
• A radiomics-to-radiology feature dictionary was used to supplement interpretability.

Main Results:

• A compact signature of three features (Textural Entropy, Calcification Index, Local Homogeneity) was identified.
• The CatBoost model achieved a macro-averaged one-vs-rest AUC of 0.87 and 66.7% accuracy.
• Local maps revealed heterogeneous texture and focal hotspots in firm tumors, linking radiomic features to spatial patterns.

Conclusions:

• The developed framework provides a proof-of-concept for interpretable radiomics in meningioma consistency prediction.
• Integrating explainability techniques enhances the link between radiomic signatures, spatial tumor patterns, and MRI characteristics.
• This approach offers a potentially extensible method for other imaging tasks influenced by tissue composition.