A subregion-based survival prediction framework for GBM via multi-sequence MRI space optimization and...

Area of Science:

• Radiology and Oncology
• Medical Imaging Analysis
• Computational Biology

Background:

• Glioblastoma (GBM) survival prediction post-radiation therapy is challenging.
• High-dimensional radiomic features from multi-sequence MRIs often lead to overfitting.
• Accurate patient stratification is crucial for personalized GBM treatment.

Purpose of the Study:

• To develop a subregion-based survival prediction framework for Glioblastoma (GBM) patients.
• To introduce a novel feature construction method optimizing multimodal MRI data.
• To improve the accuracy of 1-year and overall survival predictions.

Main Methods:

• Developed a framework involving feature space optimization and clustering-based feature bundling.
• Extracted 680 radiomic features per subregion from multi-sequence MRIs (Pyradiomics).
• Utilized 8231 initial features, reduced to 235 effective features for model construction.

Main Results:

• Achieved AUCs of 0.998 (training) and 0.983 (testing) for 1-year survival prediction.
• Outperformed initial feature set (AUCs 0.940/0.923) with the optimized feature subset.
• Predicted overall survival with a C-index of 0.872 using an ensemble regressor.

Conclusions:

• The subregion-based framework significantly enhances GBM survival prediction accuracy.
• Optimized feature selection and construction are key for robust prediction models.
• This approach facilitates better patient stratification for personalized GBM therapy.