IMG-14. Post-Therapy Glioblastoma Multiforme (GBM) Progression Assessment using Conditional Tabular Generative Adversarial Network (CTGAN)-augmented Radiomics based Imaging Signature

Abstract BACKGROUND Differentiating pseudo-progression (PsP) from true tumour progression (TP) in patients with GBM after standard chemoradiotherapy is a significant diagnostic challenge because of the substantial overlap in their radiological characteristics. Precise differentiation is very important for optimizing therapeutic prognosis and decisions. Advanced computational methodologies have demonstrated promising results in enhancing diagnostic precision. Furthermore, Generative Adversarial Networks have emerged as powerful tools for addressing class imbalance and improving model robustness by synthetically augmenting data in limited data scenarios such as GBM progression assessment. METHODS This study utilized multiparametric structural MRI sequences (T1, T2, FLAIR, and T1GD) from 58 GBM patients which comprises 41 cases of TP and 17 cases of PsP. Tumor regions were segmented using the nnU-Net framework and regions of interest were defined for radiomic feature extraction. Handcrafted features such as shape, texture, and intensity were extracted. To mitigate class imbalance a CTGAN was implemented. Feature selection was performed using a Variational Autoencoder followed by dimensionality reduction with Principal Component Analysis. A 5-fold cross-validation strategy was employed to train and evaluate multiple imaging signatures for distinguishing between PsP and TP. RESULTS The SVM based imaging signatures yielded promising results with an accuracy of 91.84% ± 0.0359 (95% CI – 0.841 to 0.989) and AUC of 0.9667 ± 0.378 (95% CI - 0.920 to 0.998). This demonstrates the model’s strong predictive performance and robustness given the limited dataset of 58 patients. The low standard error further indicates consistent performance across folds which affirms the reliability of the classification pipeline. CONCLUSION This study demonstrates the effectiveness of a hybrid radiomics and machine learning framework enhanced with CTGAN for distinguishing PsP from TP in GBM. The SVM model achieved high accuracy with consistent performance across folds. These results highlight the potential of AI-driven radiomic models as reliable tools for post-therapy GBM assessment.