Answers are grounded in this paper and cite their sources.
Learning 3D brain lesion segmentation from bounding-box (bbox) annotations requires converting coarse spatial constraints into reliable voxel-wise supervision. We present TRUST-Seg (Teacher-Refined Uncertainty-aware Soft-label Training), a weakly supervised framework that combines heterogeneous pseudo-label generation with confidence-guided target refinement. Dataset-specific ensembles of frozen SAM-family vision foundation models and GrabCut generate reliability-weighted soft pseudo-labels, while inter-teacher disagreement weights the training of an initial 3D U-Net. The central contribution is Smooth Delta Attention (SDA), which computes a voxel-wise blending gate from the difference between student self-confidence and teacher-ensemble certainty. A student-confidence safeguard and bbox constraints regulate refinement, and a second 3D U-Net is trained from scratch on the refined targets. Both students receive MRI and bbox inputs during training and inference. Neither uses voxel-level masks as training targets, but dense validation masks calibrate teacher reliability and support model selection. Under an evaluation protocol using tight slice-wise boxes simulated from reference masks, TRUST-Seg achieves DSC scores of 85.47% on ISLES 2022 and 91.11% on BraTS 2019, exceeding the strongest evaluated weakly supervised methods by 3.26 and 0.38 Dice points, respectively. The corresponding gaps to matched fully supervised references (86.85% and 91.97%) are 1.38 and 0.86 Dice points. These results support confidence-guided pseudo-label refinement under the evaluated bbox-conditioned setting.
No takes yet. Share an insight, caveat, or question.
Alshurbaji et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: