Abstract

We describe the FLAIR-Viaevum pipeline for three segmentation tasks of the BraTS 2026 cluster of challenges: brain metastases segmentation in post-treatment MRI (BraTS-METS), pre- and post-treatment pediatric tumor segmentation (BraTS-PEDs), and segmentation generalizability across tumor types (BraTS-GoAT). All three tasks are scored with region-wise overlap and boundary metrics in which a single spurious connected component can zero out the lesion-wise score of a region, and every task imposes a hard limit of one scored submission per day per team. Under this budget, the central engineering question is not only does a mechanism help'' butcan we know it helps before spending a leaderboard slot on it’’. Our answer is a two-stage transfer gate: (a) an offline cross-validation harness that reproduces the official scoring rules exactly, including lesion-wise matching, the 50-voxel suppression of small predicted components and the treatment of regions absent from the ground truth; and (b) a firing-volume parity check on the official validation inputs, because any mechanism that changes how much a model fires interacts with lesion-wise division in ways that cross-validation cannot see. Gated by this procedure we combine nnU-Net-family backbones ensembled across architectures and seeds with task-specific, metric-aware post-processing: seeded hysteresis region growing with top-1 resection-cavity selection for BraTS-METS, dual-seed ensembling with calibrated minority-class probability scaling and a cystic-component specialist for BraTS-PEDs, and three-architecture ensembling with small-component removal of enhancing-tumor predictions for BraTS-GoAT. On the public validation leaderboards the protected submissions reached simple proxy averages of 0.554 over the twelve published BraTS-METS metrics (rank 22 of 70), 0.526 over the twelve BraTS-PEDs metrics (rank 6 of 32) and 0.670 over the six published BraTS-GoAT metrics (rank 17 of 47) at the July 23, 2026 snapshot. Equally important, the gate rejected nine mechanisms whose cross-validation gains would not have transferred, including one with a +0.092 cross-validation gain that inverted to -0.008 on the leaderboard. We report the pipeline, the gate, and the negative results that shaped it.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_PEDs_014.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to Open Review

Open Review Page: https://openreview.net/forum?id=1XLT8bEna1

BibTex

@InProceedings{LaiLon_MechanismAware_MICCAISAT2026,
        author = { Lai, Long AND Zhang, Yuling AND Lin, Hansheng AND Hu, Minli AND Liu, Wanwan AND Ke, Chao},
        title = { { Mechanism-Aware Post-Processing Under a Submission Budget: FLAIR-Viaevum at the BraTS 2026 METS, PEDs and GoAT Challenges } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17253},
        month = {pending},
        page = {pending}
}


back to top