Back to all papers

Machine learning models for predicting pseudoprogression in glioblastoma: A systematic review and diagnostic meta-analysis.

June 12, 2026pubmed logopapers

Authors

Kotochinsky M,Silva YP,Goyal A,Łajczak P,Rios-Garcia W,Suruagy-Motta R,Petry INS,Fagundes TP,Bertani R

Affiliations (10)

  • Department of Medicine, National University of Cuyo, Mendoza, Argentina.
  • Healthcare Institute of South Iceland, Selfoss, Iceland.
  • Department of Neurology, JFK University Medical Center, Edison, NJ, United States.
  • Department of Medicine, Medical University of Silesia, Katowice, Poland.
  • Department of Medicine, National University of San Luis Gonzaga, Ica, Peru.
  • Department of Medicine, Cesmac University Center, Maceio, Brazil.
  • Department of Medicine, University of Southern Santa Catarina, Palhoca, Brazil.
  • Department of Head and Neck Surgery, Barretos Cancer Hospital, Barretos, Brazil.
  • Hospital Israelita Albert Einstein-Unidade Alphaville, Barueri, Brazil.
  • Department of Neurosurgery, University of São Paulo, Sao Paulo, Brazil.

Abstract

Artificial intelligence and machine learning (ML) are increasingly being used in medical practice, particularly in neuroradiology. Glioblastoma, a highly aggressive brain tumor, poses significant diagnostic challenges, especially in differentiating actual tumor progression (TP) from pseudoprogression (PsP) using conventional magnetic resonance imaging (MRI). Machine learning and AI techniques offer a promising approach, but their diagnostic accuracy remains debated. A systematic review and diagnostic meta-analysis were performed to assess the diagnostic efficacy of ML algorithms applied in MRI of glioblastoma patients to differentiate between TP and PsP. We evaluated classical ML and deep learning models using diffusion-weighted, perfusion-weighted, and structural MRI inputs to differentiate PsP from true progression in glioblastoma. PubMed, Scopus, and Cochrane databases were searched. Pooled sensitivity, specificity, diagnostic odds ratio (DOR), accuracy, and area under the curve (AUC) were calculated using a random-effects model. Eleven studies involving a total of 676 patients were included. Machine learning models demonstrated a pooled sensitivity of 0.899 (95% CI, 0.843-0.936; <i>I</i>² = 16.3%) and a specificity of 0.766 (95% CI, 0.647-0.854; <i>I</i>² = 38.8%). The DOR was 27.059 (95% CI, 11.094-65.999; <i>I</i>² = 61.9%), and the overall accuracy was 0.85 (95% CI, 0.788-0.897; <i>I</i>² = 66%). The AUC was estimated at 0.88 (95% CI, 0.727-0.896). Deek's funnel plot showed no significant publication bias (<i>P</i> = .283). Machine learning-based models demonstrate high diagnostic accuracy in distinguishing TP from PsP in glioblastoma patients, offering a promising tool to enhance radiological assessment and support clinical decision-making. However, further standardization across algorithms and datasets is needed to facilitate their integration into routine clinical practice.

Topics

Journal ArticleReview

Ready to Sharpen Your Edge?

Subscribe to join 11k+ peers who rely on RadAISlice. Get the essential weekly briefing that empowers you to navigate the future of radiology.

We respect your privacy. Unsubscribe at any time.