Machine learning for grading prediction and survival analysis in high grade glioma

Area of Science:

• Radiology and Medical Imaging
• Oncology
• Machine Learning in Medicine

Background:

• High-grade glioma (HGG) classification is critical for treatment planning.
• Accurate differentiation between glioma grades (e.g., Grade III vs. Grade IV) remains a challenge.
• Magnetic resonance imaging (MRI) offers rich data for quantitative analysis.

Purpose of the Study:

• To develop and validate an MRI-based radiomics model for HGG classification.
• To identify the optimal machine learning (ML) approach for differentiating HGG grades.
• To evaluate the performance of various ML models in classifying HGG.

Main Methods:

• Retrospective analysis of 184 HGG patients (59 Grade III, 125 Grade IV).
• Extraction of radiomics features from T1-weighted MRI (T1WI).
• Application of LASSO for feature selection and seven ML classifiers (Logistic Regression, XGBoost, Decision Tree, Random Forest, Adaboost, Gradient Boosting, Stacking fusion).
• SMOTE technique used to address data imbalance.

Main Results:

• XGBoost classifier showed the best performance among non-fusion models.
• SMOTE improved classifier performance by addressing data imbalance.
• The Stacking fusion model achieved the highest performance with an AUC of 0.95, sensitivity of 0.84, accuracy of 0.85, and F1 score of 0.85.

Conclusions:

• MRI-based quantitative radiomics features demonstrate strong performance in HGG classification.
• The XGBoost model is superior to other non-fusion classifiers.
• The Stacking fusion model significantly outperforms non-fusion approaches for HGG grading.