This AI study analyses multiparametric MRI from 58 glioblastoma patients to distinguish true progression from pseudoprogression, comparing handcrafted radiomics, Vision Transformer features and hybrids while using CT-GAN for class imbalance.
Key findings
- The radiomics SVM performed best, with mean accuracy 91.84%±3.59 and AUC 0.9667±0.0378 within cross-validation. The result suggests quantitative image features may help distinguish these radiologically overlapping states.
Why this matters globally
If validated across hospitals, scanners and populations, the system could reduce inappropriate treatment changes, but it should support multidisciplinary assessment rather than replace diagnosis.
Thai researcher contribution
Navneet Kumar Dubey, jointly affiliated with Shinawatra University and overseas institutions, contributed to the AI framework. The accessible record does not establish that the patient cohort was Thai.
Limitations to consider
With 58 patients and 17 PsP cases, overfitting risk is high. GAN generation, dimensionality reduction and feature selection must occur within each fold to avoid leakage, which the abstract does not confirm. There is no external validation, calibration, decision curve, prospective test or scanner-shift analysis; synthetic data do not replace new patients.