Optimizing machine learning-based multimodal radiomics for predicting IDH status in gliomas: A SHAP-based multicenter...

Area of Science:

• Neuroimaging
• Machine Learning in Oncology
• Radiomics

Background:

• Predicting isocitrate dehydrogenase (IDH) status is critical for glioma management.
• Conventional MRI (cMRI) has limitations; radiomics and machine learning (ML) face challenges with single-center data and interpretability.
• There is a need for interpretable, multicenter ML models integrating diverse MRI sequences.

Purpose of the Study:

• To develop and validate an interpretable, multicenter ML model for predicting IDH status in gliomas.
• To integrate conventional MRI (cMRI) with functional sequences (DWI and PWI) for enhanced prediction.
• To improve the clinical translation of ML models in neuro-oncology.

Main Methods:

• Retrospective study with 180 patients from four institutions (150 training, 30 external test).
• Radiomics features extracted from cMRI, DWI, and DSC-PWI (CBV maps).
• Multiparametric MRI fusion radiomics models built using ML algorithms, with SHapley Additive exPlanation (SHAP) for interpretability.

Main Results:

• The full-modality model (cMRI + DWI + PWI) with 3D modified segmentation achieved AUC of 0.840 (training) and 0.810 (external test).
• Functional sequences significantly improved prediction accuracy compared to cMRI alone.
• SHAP analysis identified key predictive features and provided individualized model explanations.

Conclusions:

• The developed ML-SHAP model reliably predicts IDH status and demonstrates multicenter generalizability.
• This interpretable tool integrates conventional and functional MRI for glioma management.
• The model shows potential for supporting preoperative molecular diagnosis in gliomas.