GlioMIL: A Frozen-Backbone Multiple-Instance Learning Framework for Calibrated and Explainable Glioma Radiogenomics

Abstract Non-invasive prediction of glioma molecular status from magnetic resonance imaging (MRI) could reduce reliance on invasive biopsy. Still, deep-learning pipelines for this task are often trained on small cohorts and rarely assessed for calibration or explainability. We present GlioMIL, a 2.5D multiple-instance-learning (MIL) framework predicting isocitrate dehydrogenase (IDH) mutation status, tumour grade, O 6 -methylguanine-DNA methyltransferase (MGMT) promoter methylation, and 1p/19q codeletion from multi-contrast MRI, representing each tumour as a bag of tumour-guided axial slices embedded by a frozen, ImageNet-pretrained backbone and aggregated via attention-based pooling feeding multi-task heads. On a publicly available glioma cohort, a controlled five-way ablation with bootstrap confidence intervals and paired significance testing shows this frozen-backbone approach matches a fully-supervised 3D convolutional neural network (CNN) baseline, and that fusing standardised clinical and radiomic features into the pooled representation significantly improves over image-only multi-task learning for IDH and grade. However, a plain classifier on clinical and radiomic features alone, with no image input, matches or significantly exceeds the fused model on these tasks, so the frozen image branch adds no predictive value beyond structured features. MGMT and 1p/19q codeletion remain close to chance across all conditions. The fusion model is label-efficient and reasonably well calibrated, but its attention weights correspond to tumour regions only at a near-chance level, cautioning against treating attention as a reliable spatial explanation. We report our evaluation protocol, including rejection of an earlier flawed ensembling scheme and a per-fold feature-normalisation correction, to support trustworthy, reproducible reporting.