What question did this study set out to answer?

This research aims to differentiate true progression from pseudoprogression in glioblastoma multiforme post-therapy using advanced AI techniques.

July 8, 2026Open Access

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

Key Points

This research aims to differentiate true progression from pseudoprogression in glioblastoma multiforme post-therapy using advanced AI techniques.
Analyzed multiparametric MRI data from 58 GBM patients.
Utilized feature extraction strategies including handcrafted radiomics and deep learning features.
Addressed class imbalance with Conditional Tabular Generative Adversarial Network and applied PCA to reduce dimensionality.
The SVM model achieved an accuracy of 91.84% ± 3.59 with a 95% CI of 0.841 to 0.989.
The model's AUC was 0.9667 ± 0.0378 with a 95% CI of 0.920 to 0.998, indicating strong performance.
Results support the clinical applicability of the CT-GAN-augmented radiomics framework for distinguishing progression types.

Abstract

Objective: Differentiating true progression (TP) from pseudoprogression (PsP) in post-therapy of glioblastoma multiforme (GBM) remains a critical challenge in neuro-oncology, due to their overlapping radiological characteristics. Consequently, an accurate, non-invasive differentiation between PsP and TP is vital for informing treatment decisions and improving patient outcomes. Therefore, we aimed to present a comprehensive AI-driven framework integrating handcrafted radiomics, deep learning features, and generative modelling to enhance diagnostic precision. Methods: Multiparametric MRI data (T1, T2, FLAIR, T1GD) from 58 GBM patients (41 TP, 17 PsP) were analysed using three feature extraction strategies: handcrafted radiomics, Vision Transformer (ViT)-based deep features, and a hybrid combination of both. To address class imbalance, synthetic samples were generated using a Conditional Tabular Generative Adversarial Network (CT-GAN). Feature dimensionality was reduced through a two-stage pipeline comprising a Variational Autoencoder (VAE) followed by Principal Component Analysis (PCA), preserving 95% variance. Classifiers, including Support Vector Machine (SVM), Random Forest, XGBoost, and Multi-Layer Perceptron (MLP), were trained and validated via stratified 5-fold cross-validation. Results: The radiomics-based SVM model yielded the best performance with an accuracy of 91.84% ± 3.59 (95% CI - 0.841 to 0.989) and an AUC of 0.9667 ± 0.0378 (95% CI - 0.920 to 0.998), demonstrating promising performance within the studied dataset and highlighting potential for clinical application, subject to further external validation. Conclusion: Our demonstrated CT-GAN-augmented hybrid radiomics framework offers a robust, non-invasive, and promising solution for reliably distinguishing PsP from TP. It shows promising performance within the studied dataset, with potential for future clinical translation subject to external validation.

اسأل الذكاء الاصطناعي

Bookmark

View Full Paper