Abstract
We propose DeJEPA, a Dense Joint-Embedding Predictive Architecture for non-contrastive self-supervised representation learning in 3D medical images. DeJEPA adapts joint-embedding predictive learning to volumetric data, combining a voxel-level invariance loss with Sketched Isotropic Gaussian Regularization, a distribution-matching regularizer that analytically aligns the embedding space with an isotropic Gaussian distribution. This principled objective enforces informative, decor-related representations without negative pairs. To reduce within-batch dependence for this objective, we further introduce an inter-patient sampling strategy paired with intensity-focused augmentations that promote sample diversity and sufficient view variance across the batch. Built on a multi-scale 3D feature pyramid backbone, our framework learns rich voxel-level features directly from unlabelled 3D medical images. Experiments performed on the BTCV and MSD public datasets demonstrate that DeJEPA consistently outperforms comparative self-supervised learning (SSL) baselines on probing experiments, achieving a 66.4% linear and 74.1% nonlinear probing Dice score on BTCV and improving 1-shot finetuning by 24.5 Dice points over training from scratch. These results highlight the potential of theoretically grounded, non-contrastive objectives for dense voxel-level representation learning.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MLMI_011.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=OuxfEPJ83f
BibTex
@InProceedings{PorAlo_DeJEPA_MICCAISAT2026,
author = { Portafaix, Aloys AND Kadoury, Samuel},
title = { { DeJEPA: Non-Contrastive Self-Supervised Learning for Voxel-Level Representations } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17272},
month = {pending},
page = {pending}
}
