Externally validated explainable 3D CNN ensemble model for non-invasive prediction of IDH and MGMT status in gliomas

To develop and externally validate an explainable 3D convolutional neural network (CNN) ensemble for non-invasive prediction of IDH mutation and MGMT promoter methylation status in gliomas using routine MRI.

Multi-institutional MRI data were curated from the UCSF-PDGM and UPENN-GBM cohorts for training and testing, with independent external validation on the MU-Glioma-Post dataset. Separate 3D CNNs were trained for IDH and MGMT classification, and the models were combined into an ensemble metaclassifier. Model interpretability was assessed using Integrated Gradients saliency maps for voxel-level attributions and Shapley values for modality-level contributions.

The ensemble model achieved an AUC of 0.94 (95% CI, 0.84-1.00) in testing and 0.77 (95% CI, 0.71-0.82) in external validation for IDH mutation classification. The ensemble yielded AUCs of 0.81 (95% CI, 0.64-0.94), 0.68 (95% CI, 0.47-0.88), and 0.57 (95% CI, 0.52-0.63) for MGMT promoter methylation in validation, testing, and external validation, respectively. Integrated Gradients (IG) maps highlighted biologically plausible regions, with IDH predictions focusing on peritumoral FLAIR hyperintensity and MGMT predictions emphasizing the enhancing tumor core on T1CE. Shapley analysis showed FLAIR contributed 64% and T1CE 36% to IDH classification, while MGMT classification relied more heavily on T1CE (60.6%), with complementary input from FLAIR (30.8%).

We present an explainable 3D ensemble deep learning framework capable of predicting IDH mutation status from routine MRI with moderate external generalizability. While the same framework achieved strong internal performance for MGMT methylation status, this did not generalize on external validation, underscoring that MGMT prediction from imaging alone remains an unresolved challenge requiring further methodological development.