Abstract
In this challenge, we apply a single pipeline built on the Medical Imaging Segmentation Toolkit (MIST) across three distinct tasks – brain metastases, pediatric tumors, and cross-population generalizability. To determine our optimal submission under compute constraints, we evaluate two competing deployment strategies. The first is our returning from-scratch pocket nnUNet baseline, a two-time third-place finisher. The second is a preliminary application of the Medical Imaging Semantic Foundation Toolkit (MISFIT), an open-source framework that we are actively developing for masked autoencoder pretraining of 3D medical imaging encoders. We use MISFIT to pretrain a SwinUNETR-V2 encoder on a pooled, cross-task corpus of over 16,000 unlabeled MRI volumes. Fine-tuned on each task’s labeled cohort, the MISFIT-pretrained models reach their loss plateau well inside their reduced training budget and are statistically favored on two of three tasks under cross-validation. Every one of those advantages vanishes in held-out validation, which also reveals a significant regression in brain metastases. We show that pretraining on the training split inflates cross-validation in the pretrained model’s favor, making held-out scoring the only protocol we trust for this decision. Because our two arms differ in architecture, patch size, fine-tuning budget, and initialization, this comparison evaluates two distinct system-level pipelines rather than isolating the effect of pretraining. Our final submissions for all three tasks use the baseline model: the challenge rules do not allow pooling data from other sub-challenges for the final submission, and this decision is independently supported by the lack of a defensible held-out improvement from pretraining.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_GoAT_005.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/BraTS_GoAT_005_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=UO1d2SOKmt
BibTex
@InProceedings{CelAdr_MISFIT_MICCAISAT2026,
author = { Celaya, Adrian AND Kumar, Yogesh AND Twam, Awj AND Riviere, Beatrice AND Fuentes, David AND Netherton, Tucker},
title = { { MISFIT: A Self-Supervised Pretraining Toolkit for 3D Medical Imaging — A Preliminary Test Across Three BraTS 2026 Segmentation Tasks } },
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}
}