GAN-Augmented Radiomics and Machine Learning for Post-Therapy GBM Progression Assessment

Area of Science:

• Neuro-oncology
• Medical Imaging Analysis
• Artificial Intelligence in Medicine

Background:

• Distinguishing true progression (TP) from pseudoprogression (PsP) in glioblastoma multiforme (GBM) post-therapy is challenging due to similar radiological features.
• Accurate, non-invasive differentiation is crucial for effective treatment planning and improved patient outcomes in neuro-oncology.

Purpose of the Study:

• To develop and validate a comprehensive AI-driven framework for precise differentiation between TP and PsP in GBM patients.
• To integrate handcrafted radiomics, deep learning features, and generative modeling for enhanced diagnostic accuracy.

Main Methods:

• Analysis of multiparametric MRI data from 58 GBM patients using radiomics, Vision Transformer (ViT)-based deep features, and hybrid approaches.
• Utilized Conditional Tabular Generative Adversarial Network (CT-GAN) for synthetic data generation to address class imbalance.
• Implemented a two-stage dimensionality reduction (VAE and PCA) followed by classifier training (SVM, Random Forest, XGBoost, MLP) and cross-validation.

Main Results:

• The radiomics-based Support Vector Machine (SVM) model achieved the highest performance.
• Achieved an accuracy of 91.84% ± 3.59 and an Area Under the Curve (AUC) of 0.9667 ± 0.0378.
• Demonstrated promising performance for potential clinical application, pending external validation.

Conclusions:

• The CT-GAN-augmented hybrid radiomics framework provides a robust, non-invasive method for distinguishing PsP from TP in GBM.
• The study highlights the potential for AI in improving diagnostic precision for post-therapy GBM assessment.
• Further external validation is recommended for clinical translation of this promising approach.