Abstract

We ask whether, for a fixed containerized segmentation system, a commit-frozen multi-gate policy audit can prevent adoption of an inference modification whose aggregate development-subset benefit does not transfer consistently to a same-corpus policy-selection holdout. GAT-26 is a five-fold region-based residual-encoder nnU-Net trained only on subchallenge data from a recorded random initialization. Our contribution is procedural: candidate inference modifications were executed end-to-end through the release machinery on excluded-fold predictions, with candidates, utility and gates fixed in a pushed commit before the specified comparison. The utility is the unweighted mean of six raw component means, ET/TC/WT by DSC and NSD, on common subject support. Eightfold mirroring improved all six DSC/NSD components on the development folds but failed a frozen lesion false-negative veto and stopped. A separately commit-frozen follow-up opened the untouched holdout folds; the gain was descriptively smaller: +0.00073 against +0.00253 at the official-ranking tolerance, +0.00247 against +0.00417 in a lower-tolerance sensitivity analysis. The holdout bootstrap intervals spanned zero across tolerances, one fold changed sign, and the reference-lesion miss rate regressed by +0.0070 against a 0.0025 operational margin, so the baseline was retained. We document a discrepancy easy to ship unnoticed: our cross-validation used eightfold mirroring while the released container does not, so they are not comparable. This is a bounded case study, not evidence the procedure is preferable to nested validation.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_GoAT_021.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/BraTS_GoAT_021_supp.pdf

Link to Open Review

Open Review Page: https://openreview.net/profile?id=%7ENathan_Chen1

BibTex

@InProceedings{CheNat_GAT26_MICCAISAT2026,
        author = { Chen, Nathan},
        title = { { GAT-26: Release-Path Auditing and Confirmation-Gated Inference Selection for Cross-Tumor Brain Tumor Segmentation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17254},
        month = {pending},
        page = {pending}
}


back to top