A Stability Audit for Resource-Efficient Brain Tumor MRI Segmentation

Shorter magnetic resonance imaging (MRI) protocols could reduce scanner occupancy where hospitals have limited imaging capacity, but a sequence subset selected from one training run may fail under another model or patient partition. We present a protocol-selection stability audit for abbreviated brain tumor MRI segmentation. The audit first evaluates all 15 non-empty subsets of T1, contrast-enhanced T1 (T1c), T2, and fluid-attenuated inversion recovery using six shared modality-dropout models. It then trains 20 dedicated models for the two leading two-sequence candidates across five repeated source partitions and two architectures. Each partition separates checkpoint selection, protocol selection, and final testing. Shared models favor T1c+FLAIR externally in all six runs, suggesting consensus. Dedicated models dissolve that consensus: three of ten source selections reverse on final testing, external data agree with the selected protocol in four of ten comparisons, and statistically resolved external effects split evenly between candidates. The result reframes abbreviated-protocol design as a stability problem. This retrospective audit evaluates segmentation reliability; diagnostic equivalence and deployable protocols remain outside its scope. Both cohorts come from high-resource clinical settings.