FOMO26

MICCAI Logo

Schedule

Half-day event at MICCAI 2026

October 1, 2026 | 13:30 - 18:00 | Strasbourg, France

Speakers

Program

Location: Auditorium Cassin (G), Strasbourg Convention Centre
13:30 - 13:40

Opening remarks

13:40 - 14:25

Keynote 1

Randall Balestriero (Brown University)

Counterfactual JEPAs for World Models in the Real World

14:25 - 14:50

Challenge design and motivation

14:50 - 15:30

Presentations from participants (method track)

  • BAIA-BrainFM
  • BrainAI
  • Azmuth
Break
16:00 - 16:45

Keynote 2

Corentin Dancette (Raidium)

Foundation models in radiology: One model for all tasks?

16:45 - 17:15

Presentations from participants (open track):

  • CGVision
  • bmic
  • FariaTeam
17:15 - 18:00

Results and award ceremony

Talk by Randall Balestriero

Abstract: Counterfactual JEPAs for World Models in the Real World

This talk presents a sequence of recent results on Joint-Embedding Predictive Architectures (JEPAs), spanning self-supervised representation learning, action-conditioned world models, and counterfactual reasoning. We begin with LeJEPA, which addresses representation collapse through a distributional principle: embeddings are regularized toward an isotropic Gaussian using Sketched Isotropic Gaussian Regularization, producing a simple predictive objective with theoretical guarantees and without teacher–student networks or stop-gradient heuristics. We then introduce LeWorldModel, which extends this construction to sequential data by learning an encoder and action-conditioned latent transition model, end-to-end from raw observations. By combining next-embedding prediction with Gaussian latent regularization, the model supports planning and control directly in representation space, without reconstructing future pixels. Finally, we discuss new directions in counterfactual JEPAs, where object-level masking and latent interventions create structured partial-observability problems that force the predictor to model interaction-dependent dynamics rather than exploit local correlations. Throughout, we emphasize the geometric and statistical principles connecting these methods, and ask when latent prediction is sufficient for learning causal, controllable abstractions of a dynamical world.

Bio

Randall is an Assistant Professor of Computer Science at Brown University, and a leading expert on self-supervised learning, JEPAs, and World Models.

Talk by Corentin Dancette

Abstract: Foundation models in radiology: One model for all tasks?

In this talk, he will review recent work on foundation models for radiology and present Raidium's current research, looking at how well existing models can support different tasks such as abnormality classification, fine-grained segmentation and report generation.

Bio

Corentin Dancette is Chief Research Officer at Raidium, where he leads research on foundation models for radiology. His work spans image-text alignment, radiology report generation, self-supervised learning, image segmentation, and object detection, and their team open-sourced foundation models such as Curia, Jolia, and RadSAM. Before joining Raidium, he completed his PhD at Sorbonne Université, where he worked on deep learning at the intersection of computer vision and language.