A logistic paradigm to predict molecular glioblastoma on clinical MRI

This study aims to externally validate a logistic strategy for predicting the integrated molecular diagnosis of glioblastoma (mGBM) as defined by the World Health Organization (WHO) 2021 Classification of Tumors of the Central Nervous System (CNS).

Ethical approval was obtained for a retrospective study of diffuse gliomas (CNS WHO grades 2-4) with available isocitrate dehydrogenase (IDH) and 1p19q status and whole genome profiling for mGBM, where appropriate. Age and MRI characteristics (apparent diffusion coefficient (ADC), tumor location, contrast enhancement pattern, cyst(s), and calcification) were scored by two observers blinded to the tissue results and to each other's results. Observer concordance was assessed using intraclass correlation coefficient (ICC) and Cohen's kappa. Diagnostic performance, comparisons using McNemar and DeLong tests, and model calibration were obtained in SPSS 27 and Python (v3.12).

A total of 108 consecutively enrolled cases were eligible from January to December 2023 (mean age 56.9 ± 15.5 years; range 20-82), comprising 85 mGBM/IDH wild-type (IDH^wt) and 23 IDH mutant (IDH^mut) gliomas. Model sensitivity ranged from 97% to 99%, with 48%-91% specificity, 87%-97% accuracy, and AUCs of 0.93-0.98. For the trainee, model use produced increases in sensitivity from 87.1% to 97.6% with model A and 96.5% with model B and in accuracy from 84.3% to 87.0% and 91.7%, respectively. No within-observer difference in overall accuracy and no paired AUC comparison remained statistically significant after Holm adjustment. Model B demonstrated more consistent calibration across observers. Among eight mGBM lacking rim enhancement, correct classification by the trainee increased numerically from 12.5% to 100% with model A and 87.5% with model B.

The logistic strategy achieved high sensitivity and discrimination for identifying mGBM according to CNS WHO 2021. Model use produced increases in trainee performance, particularly sensitivity, while Model B demonstrated more consistent calibration across observers. These findings support the further evaluation of the tool, particularly in radiologically atypical mGBM.