Detector-Stage Failures Account for Most of the Increase in the External Deficit of a Box-Prompted Glioma Segmentation Pipeline on BraTS-Africa: A Three-Arm Decomposition Across Two External Cohorts

Background/Objectives: When a box-prompted segmentation pipeline loses accuracy on new data, the detector proposing the box or a later stage may be responsible; this study measured which. Methods: A real-time detection transformer (RT-DETR) feeding the Segment Anything Model 2.1 was compared with a detector-free model (MedSAM3 with a low-rank adapter) on whole-tumor Dice in 158 patients with glioma: held-out Brain Tumor Segmentation (BraTS) 2020 cases (n = 24), the RHUH-GBM glioblastoma cohort (n = 39), and BraTS-Africa (n = 95, glioma presumed). Giving the segmenter the ground-truth box splits the Dice gap between the systems into a detector-stage term and a residual. Results: The detector-stage term was +0.006 Dice on held-out cases, +0.018 on RHUH-GBM, and +0.091 on BraTS-Africa. Its increase over the held-out value was +0.0854 [95% CI +0.0450, +0.1303] on BraTS-Africa (Holm-adjusted p < 0.001) and, after correction of a preprocessing fault, +0.0120 [−0.0062, +0.0295] on RHUH-GBM, which was not resolved. On BraTS-Africa, it came mainly from slices without a detector box. No residual change was resolved. Conclusions: On BraTS-Africa, the detector stage accounted for most of the increase in the pipeline’s gap; on RHUH-GBM, no increase was resolved. The decomposition should be measured per cohort.