Robust Brain MRI Tumor Segmentation via Normal-reference Alignment, Label-free Distillation, and Uncertainty-gated Test-time Adaptation

Segmentation of glioma subregions in multi-modal MRI often degrades across scanners and protocols, undermining reliability.We propose a compact training-inference recipe that unifies three plug-in ideas: FAM-Lite, a normal-reference alignment block that contrasts patient representations with a learned healthy-brain prior to amplify pathology; label-free distillation that transfers boundary-aware cues from anatomy-respecting synthetic tumors via confidence-masked consistency; and uncertainty-gated test-time adaptation (TTA-UQ) that triggers a single recalibration step and a small TTA vote only when predictive entropy is high, improving calibration while updating only running normalization statistics in predefined decoder convolutional blocks at inference time.Integrated into a 3D Transformer-U-Net and trained with standard BraTS protocols, the method consistently surpasses nnU-Net and UNETR on BraTS-2021/2023, yielding higher Dice, lower Hausdorff95, and reduced Expected Calibration Error, with the largest gains on small enhancing-tumor foci and boundary fidelity.Zero-shot tests on BraTS-METS demonstrate stronger generalization under shift.Under a unified efficiency accounting (model parameters, GFLOPs, and latency measured under the fixed patch/overlap inference setting used throughout the paper), the recipe incurs ~5-7% parameter overhead, with modest compute and latency increases.The recipe is simple to adopt through curriculum training and plug-in modules, offering a practical path to calibration-aware, shift-robust neuro-oncology segmentation with fully specified split protocol, preprocessing, hyperparameters, and TTA-UQ settings to support reproducibility.