Evidential Safety-Gated Segmentation for Robust Brain Tumor Analysis with Missing MRI Modalities

Accurate brain tumor segmentation from multi-parametric MRI is indispensable for clinical decision-making, yet real-world deployment is persistently challenged by missing or corrupted modalities arising from scanner failures, protocol variations, and patient-specific contraindications. Existing models tend to fail silently under such conditions, propagating high-confidence but clinically incorrect predictions into downstream workflows. We present Fail-SafeAI, a safety-gated, uncertainty-driven framework that integrates evidential deep learning with a learnable clinical safety gate for reliable segmentation under modality absence. The backbone operates on the two most discriminative MRI channels — FLAIR and contrast-enhanced T1 (T1ce) — and is trained with aggressive per-channel dropout to build inherent robustness. Network outputs are parameterized as a Dirichlet distribution over four semantic classes, enabling direct epistemic uncertainty (vacuity) estimation in a single forward pass. At inference, the clinical safety gate evaluates per-slice uncertainty statistics and dynamically decides whether to accept the prediction, escalate to a radiologist, or trigger a lightweight conditional U-Net inpainter for missing channel synthesis. On BraTS 2020, Fail-SafeAI achieves a whole-tumor Dice Similarity Coefficient (DSC) of 0.864 under complete data and ≥0.795 under single-modality dropout, outperforming AdaMM by 7.2 percentage points under missing T1ce while reducing the Unsafe False Acceptance Rate from 5.1% to 0.2%. Uncertainty maps are demonstrably calibrated: vacuity is statistically higher for incorrect predictions than for correct ones, and the accuracy-rejection curve shows that deferring the 30% most uncertain pixels to a human expert raises retained accuracy from 68.8%