SYNAPSE-X: a calibration-aware vision transformer framework for selective referral in multi-class brain tumor MRI classification

Abstract

Brain tumor MRI classification models can achieve high aggregate accuracy, but reliable decision support also requires calibrated confidence estimates, transparent handling of uncertain cases, and careful evaluation under external data shift. This study presents SYNAPSE-X, a calibration-aware Vision Transformer framework for multi-class brain tumor MRI classification with uncertainty-guided selective referral. The framework classifies contrast-enhanced T1-weighted MRI images into glioma, meningioma, pituitary tumor, and no-tumor categories while explicitly evaluating prediction reliability beyond accuracy alone. The corrected experimental cohort included 7,023 images, with 5,712 training images and 1,311 held-out test images; augmentation was applied only to the training subset to reduce leakage risk. Model performance was assessed using accuracy, macro-F1, weighted-F1, expected calibration error, Brier score, negative log-likelihood, confidence-stratified accuracy, and risk-coverage analysis. Post hoc temperature scaling was applied to improve probability calibration without changing predicted labels. On the held-out test set, SYNAPSE-X achieved 98.41% accuracy, 98.2% macro-F1, and 98.4% weighted-F1. Temperature scaling reduced expected calibration error from 0.041 to 0.018 and Brier score from 0.036 to 0.021. Using a confidence threshold of 0.85, the selective-referral strategy deferred 6.7% of cases for expert review, while retained high-confidence cases achieved 99.12% accuracy. External direct-transfer testing on an independent external brain MRI cohort of 3,264 images achieved 97.86% accuracy, 97.7% macro-F1, 97.9% weighted-F1, and an expected calibration error of 0.026 without retraining, fine-tuning, external recalibration, or threshold re-selection. These findings position SYNAPSE-X as a research-stage reliability-assessment framework for calibrated brain tumor MRI classification, supporting uncertainty-aware referral while acknowledging limitations related to public image-level datasets, unavailable patient-level identifiers, and the need for prospective clinical validation.