<span data-teams="true">Detector-Stage Failures in Box-Prompted Glioma Segmentation: A Three-Arm Analysis Across Two External Cohorts

Source abstract

Background/Objectives: Whether the detector or the segmenter fails externally is unknown in a box-prompted pipeline; an oracle-box arm separates them by prompting with the ground-truth box. Methods: A real-time detection transformer (RT-DETR) with Segment Anything Model (SAM) 2.1 was evaluated against MedSAM3 with a low-rank adaptation (LoRA) adapter on 158 glioma cases from three cohorts. Three arms decompose the gap into a detector-stage term and a residual; a fourth splits the residual. Results: At whole tumour the cost of finding the box rises from +0.006 Dice in domain to +0.091 on the Brain Tumour Segmentation (BraTS) Africa cohort, a difference of +0.085 [+0.045, +0.130] that survives a second detector seed and a second segmenter. The Río Hortega University Hospital glioblastoma cohort (RHUH-GBM) did not show a statistically resolved increase; the two external cohorts differ by +0.073 [+0.035, +0.116]. On BraTS-Africa, slices with no box carry the term, not poorly placed boxes. The in-domain baseline is not categorical: the term excludes zero under a matched prompt and on tumour core. Nine of 95 pipeline cases fall below 0.5 Dice; a post hoc coverage flag identifies seven. Conclusions: On BraTS-Africa, but not RHUH-GBM, the increase in pipeline deficit was primarily associated with the detector stage.