Abstract

Deploying high-performing brain metastasis segmentation models under strict runtime constraints is a practical obstacle in both challenge and clinical settings. The BraTS-METS 2026 protocol executes each submission as a Docker container on a single GPU with a fixed eight-hour budget for the 303-case test set, which a conventional five- fold nnU-Net ensemble exceeds. We ask whether that constraint can be met without retraining or architectural change, by jointly optimizing ensemble composition and deployment. Starting from a fixed nnU-Net v2 ResEncL model trained with five-fold cross-validation, we exhaustively evaluated all ten three-fold subsets. Because the lesion-wise evaluator requires minutes to hours per case, selection used a voxel-overlap Dice applied identically to every candidate. The selected subset (folds 1, 2 and 4) was then deployed in a persistent in-memory predictor that loads each fold once, removing repeated initialization overhead. The three-fold ensemble reached a mean voxel-wise Dice of 0.7734 against 0.7725 for the five-fold ensemble (paired Wilcoxon p=0.53; equivalence established against a pre-registered margin of 0.01 Dice, TOST p<0.001, n=567), while measured end-to-end inference fell from 21.1 to 6.4 hours, a 70% reduction that meets the official budget. Jointly optimizing ensemble design and deployment therefore allows computationally intensive segmentation models to satisfy strict runtime limits without measurable loss of accuracy. The same equivalence held under the publicly available BraTS-2025 lesion-wise evaluator on the full 567-case cohort (difference +0.003; equivalence at the pre-registered 0.02 margin, TOST p<0.001).

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_METS_007.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=lEQTeVh8I0

BibTex

@InProceedings{ElAbd_RuntimeAware_MICCAISAT2026,
        author = { El-Gaghous, Abdelrahman AND Saraya, Mohamed Sabry Fouad AND Ayyad, Sarah M.},
        title = { { Runtime-Aware Ensemble Optimization for Brain Metastasis Segmentation under Strict Deployment Constraints } },
        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