SYNEXA: A Self-Supervised Dual-View Transformer with Post-hoc Symbolic Rule Extraction for Multimodal MRI Glioma Classification

Brain tumour grading models should be evaluated not only for predictive performance but also for the quality and limitations of the evidence they provide. SYNEXA is a self-supervised dual-view transformer for HGG/LGG classification from axial and coronal MRI slabs formed from T1, T1ce, T2, and FLAIR. Slice selection is mask-free: a FLAIR/T1ce intensity heuristic chooses the axial and coronal slab centres, and segmentation masks are used only after training for localization analysis. The encoder is pretrained within each training fold using masked-patch prediction and dual-view contrastive alignment and is then fine-tuned with class-weighted focal loss using the official BraTS 2020 grade labels. Native transformer attention supplies spatial model-behaviour evidence, while a shallow decision tree is fitted after fine-tuning to training-fold [CLS] embeddings as a post-hoc symbolic surrogate; it does not influence the transformer classification pathway. Across the original 15 seed-fold SYNEXA evaluations, the descriptive means were 86.1% accuracy, 78.6% balanced accuracy, 91.3% F1-score, and 87.4% AUC. To address dependence across repeated cross-validation, the three out-of-fold probabilities for each patient were averaged before inference: the resulting patient-level estimates were 85.1% accuracy (95% bootstrap CI 81.3–88.6), 75.5% balanced accuracy (69.7–81.2), 90.7% F1-score (88.1–93.1), and 90.4% AUC (86.9–93.7). In a training-budget control experiment, ResNet-18, DenseNet-121, and ViT-B/16 were retrained for a non-smaller maximum budget of 150 supervised epochs per fold; their descriptive mean accuracies were 86.8%, 86.8%, and 86.4%, respectively, and no patient-level paired comparison with SYNEXA remained significant after Holm correction. Attention localization remained modest in absolute magnitude (Dice-top20 0.218; pointing-game 0.242), yet all attention metrics exceeded a circular-shift spatial null after patient-level testing (Holm-adjusted p = 0.0009). The post-hoc tree achieved 90.7% fidelity to the transformer but very low cross-run feature-set stability (mean Jaccard 0.005), so its rules are model-specific rather than transferable clinical criteria. Per-view slab coverage was 64.7%; patient-level accuracy fell from 92.2% when both selected views contained tumour to 77.8% when only one view contained tumour and 76.5% when neither did. BraTS 2018 was not used as an external test set because all 285 of its subjects overlapped with the BraTS 2020 cohort. These results position SYNEXA as an auditable dual-view classification framework with quantitatively evaluated post-hoc explanations, rather than as a statistically dominant or clinically validated neuro-symbolic system.