Explainable volumetric deep learning from contrast-enhanced T1-weighted MRI for preoperative IDH prediction in adult-type diffuse gliomas: a two-center study

Objective To evaluate contrast-enhanced T1-weighted MRI (CE-T1WI) models for preoperative prediction of isocitrate dehydrogenase (IDH) mutation status in adult-type diffuse gliomas and explore phenotype associations of influential latent features. Methods This retrospective study included 462 patients. Center 1 provided training ( n = 259) and internal-test ( n = 65) cohorts, and Center 2 provided external validation ( n = 138). Five input–architecture combinations and six classifiers were evaluated against an interpretable logistic benchmark combining age and Visually Accessible Rembrandt Images (VASARI) descriptors. For each architecture, the classifier with the highest observed external area under the curve (AUC) was selected for reporting. Performance was assessed using discrimination, calibration, decision curves, and 100 downstream refits with fixed deep features. Shapley additive explanations (SHAP) and phenotype association analyses examined influential latent features and their relationships with clinical and imaging phenotypes. Results Region-of-interest (ROI)-constrained three-dimensional vision Transformer (3D ViT) with a support vector machine (SVM) achieved an external AUC of 0.929 (95% CI, 0.885–0.973), accuracy of 0.848, sensitivity of 0.838, specificity of 0.857, and Brier score of 0.106. The age–VASARI benchmark achieved an external AUC of 0.782. Across downstream refits, the ROI-constrained model had a mean external AUC of 0.930 and the smallest mean training-minus-external AUC difference (0.019). The same 10 features formed the highest-importance SHAP set across cohorts, with different rank orders. Feature_546 correlated with age ( r = 0.461), postoperative Ki-67 ( r = 0.562), and enhancement quality ( r = 0.678; all adjusted p < 0.001). Conclusion ROI-constrained volumetric learning from CE-T1WI showed promising cross-center IDH discrimination using one routine MRI sequence. Influential latent features were associated with postoperative Ki-67 and recognizable imaging phenotypes. SHAP attribution and these exploratory associations provide complementary perspectives on model behavior in preoperative IDH assessment.