Abstract
Foundation models promise to transfer broadly across clinical imaging tasks, yet their behavior on 3D brain MRI under realistic few-shot and out-of-domain conditions remains under-explored. A
single reused encoder must also serve tasks that demand different information: dense, spatially resolved features for segmentation, and a globally coherent embedding for classification and frozen linear probing. We present our submission to the MICCAI Foundation Model 2026
(FOMO26) Challenge Methods Track, in which pretraining is restricted to the provided FOMO300K training set of 306,207 brain MRI sequences. We adapt the DINOv2/iBOT self-supervised objective to native 3D, pairing image-level self-distillation on the classification (CLS) token with a masked patch-level objective so that one backbone supplies both signals, and pretrain two Vision Transformers (ViT) of increasing capacity, FOMO-B (ViT-B, 88.3M-parameter backbone, 963 input) and FOMO
L (ViT-L, 306.9M-parameter backbone, 1123 input), directly on single sequence. Following the challenge rules we submit a single pretrained checkpoint to the 7 proposed downstream tasks, spanning classification, regression, segmentation, and representation quality. On the hidden challenge validation leaderboard (Methods-Track leaderboard snap, 2026-07-31), our FOMO-B submission ranks 1st on meningioma segmentation (DSC=0.177, NSD=0.127), while performing comparable with top ranking teams across the classification, regression, and frozen-embedding tasks. Further to reporting internal few-shot evaluation, hidden-cohort
results, and quantifying the generalization gap under domain and scanner shift, we discuss why instance-discriminative pretraining transfers strongly to global reognition but requires longer, foreground-aware fine- tuning for fine-structure segmentation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BrainWorks_034.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/profile?id=~Spyridon_Bakas3
BibTex
@InProceedings{InnShu_Scaling_MICCAISAT2026,
author = { Innani, Shubham AND Pitarch-Abaigar, Carla AND Bakas, Spyridon},
title = { { Scaling Native 3D Self-Supervised DINOv2 for Brain MRI Foundation Models } },
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}
}