STRESS TESTING SPATIAL ROUTING OF SYNTHESISED T1CE FOR MISSING-MODALITY BRAIN-TUMOUR SEGMENTATION USING TWO-DIMENSIONAL PROCEDURAL PHANTOMS

Apparent gains from adaptive fusion of synthesised contrast-enhanced T1-weighted (T1ce) MRI may reflect unequal label budgets, global calibration, or meta-model flexibility rather than genuinely spatial routing. We stress-tested this mechanism in two separately locked 16-seed studies using procedural phantoms and a 32-pair counterfactual control. Study 1 compared a restricted spatial router (QGate-MRI-R) with ungated synthesis-assisted segmentation, label-matched controls, and balanced logistic stacking. Study 2 introduced factorised acquisition shifts, alternative stackers, case-mean and histogram-preserving shuffled weights, and a target-only exposure test. In Study 1, QGate reduced in-distribution macro Brier loss versus ungated synthesis by 2.56 × 10⁻⁵ (Holm-adjusted p = 0.0269), but balanced logistic stacking performed better in distribution and synthesis with 20 additional labels achieved lower shifted Brier loss. None of the six Study 2 comparisons was significant after Holm adjustment; both locality tests had adjusted p = 0.0551. Across-seed variation in mean regional weights greatly exceeded within-map spatial variation, although the design did not separate fitting variability from regenerated case composition. Target-only exposure reduced PSNR by 12.22 dB while all 640 case-level invariance checks passed, and identical-observable pairs established construction-specific bounds of 0.25 for ET Brier loss on the ambiguity support and 2/3 for paired ET Dice. These findings support a mechanism stress test, not robust spatial-routing or clinical claims.