Diagnostic Accuracy of Radiomics and Deep Learning for Differentiating Pseudoprogression from True Tumour Recurrence in Glioblastoma: A Systematic Review and Meta-analysis
Isabela Sacco Camara *
Department of Medicine, Metropolitan University of Santos (UNIMES), Santos, São Paulo, Brazil.
Aline Pelosini Gomes
Department of Medicine, Metropolitan University of Santos (UNIMES), Santos, São Paulo, Brazil.
Olívia Voelzke Passarin
Department of Medicine, Metropolitan University of Santos (UNIMES), Santos, São Paulo, Brazil.
Amanda Scaff Mendes
Department of Medicine, Metropolitan University of Santos (UNIMES), Santos, São Paulo, Brazil.
Kalil Sallum Haddad Dib
Department of Medicine, Metropolitan University of Santos (UNIMES), Santos, São Paulo, Brazil.
Juliano dos Santos
Oncology, National Cancer Institute (INCA), Rio de Janeiro, Brazil.
*Author to whom correspondence should be addressed.
Abstract
Background: Glioblastoma is an aggressive primary brain tumour with poor survival outcomes, where distinguishing true tumour recurrence from treatment-related pseudoprogression on post-therapy MRI remains a major clinical challenge.
Aims: The primary objective of this study was to evaluate and synthesise the diagnostic accuracy of artificial intelligence models, specifically Radiomics and Deep Learning, in differentiating pseudoprogression from true tumour recurrence in glioblastoma patients following standard treatment.
Study Design: Systematic review and meta-analysis of diagnostic test accuracy.
Methodology: Following PRISMA 2020 guidelines, we identified 10 original studies involving a total cohort of 1,102 subjects (552 in the experimental recurrence group and 550 in the pseudoprogression control group). Technical performance was assessed across three outcomes: global accuracy (Area Under the Curve), clinical performance (Sensitivity and Specificity), and diagnostic efficiency (Diagnostic Odds Ratio). Quantitative synthesis was performed using a random-effects model based on the DerSimonian-Laird method to account for clinical and technical heterogeneity.
Results: The analysis of global accuracy demonstrated absolute diagnostic stability with a pooled Area Under the Curve (AUC) of 0.86 (95% Confidence Interval: 0.82 to 0.90) and null heterogeneity (I-squared = 0.0%, P = 1.0000). While clinical performance showed significant dispersion (I-squared = 98.0%, P < .0001), diagnostic efficiency was highly significant, yielding a pooled standardised mean difference of 8.18 (95% Confidence Interval: 6.03 to 10.32; P < .0001). Deep Learning models, particularly those incorporating multimodal pre- and postoperative imaging, exhibited superior specificity (up to 94%) compared to traditional radiomics.
Conclusion: Artificial intelligence models provide robust and consistent diagnostic performance in differentiating glioblastoma recurrence from pseudoprogression. The high Diagnostic Odds Ratio supports the integration of these computational tools into clinical workflows to assist in treatment decisions and mitigate unnecessary surgical interventions.
Keywords: Glioblastoma, radiomics, deep learning, Pseudoprogression, meta-analysis, diagnostic accuracy