FOMO26 FariaTeam — brain-MRI foundation encoder weights

Pretrained encoder weights for resenc_unet_b (~90.3M parameters), the encoders behind our entries to the FOMO26 Foundation Model Challenge for Brain MRI (MICCAI 2026).

• Code, recipes and inference wrappers: https://github.com/cliu238/fomo26-fariateam
• The organizers' published results: https://fomo26.github.io/results/

Chunyi (Eric) Liu, Joshua Shun Liu, Andreia V. Faria — Johns Hopkins University.

Every number about how these encoders performed lives in the code repository, next to the evidence it came from. Nothing is restated here.

1. What this is

Five 3D convolutional encoders with identical architecture (resenc_unet_b, a residual-encoder U-Net encoder; 448 tensors per file), differing only in how they were pretrained. They are the starting weights for the task-adaptation recipe published in the code repository, and they are what you want if you intend to finetune a brain-MRI model rather than reproduce our pretraining.

They are encoder-only: the pretraining decoder and the LeJEPA projector have been removed. See §5 for why, and why it costs nothing.

2. Lineage

The two random-initialization files are not filler. Any claim that pretraining transferred has to be measured against an untrained encoder of the same architecture, run through the same recipe; publishing the controls alongside the pretrained arms is what makes that check possible for someone else.

3. Track attribution

We fielded two different foundation models, one per track, and the track assignment changed late in the challenge. Stated precisely:

The models we submitted on the Methods track never saw private data.

Methods track: a second, independently pretrained encoder that adds a cross-modal content/style objective, pretrained on FOMO300K only. It is not included in this release and will be added later.

Open track: the private-continuation encoder released here.

4. The private continuation

privcont continued pretraining for 2,133 steps on ~13,300 brain-MRI scans from a private Johns Hopkins cohort plus a replayed FOMO300K fraction.

That sentence is the entire description of that data, and it is the same description given in our filed challenge submission form. The images never left Johns Hopkins. Nothing derived from them — no images, no labels, no per-case outputs, no features — is redistributed here or anywhere else. The step count above is checkable against the file itself: privcont records global_step 24,885 against v3's 22,752.

5. Encoder-only, and why

Each published file contains state_dict (encoder tensors only) plus epoch, global_step and pytorch-lightning_version. Removed: the pretraining U-Net decoder, the LeJEPA projector and SIGReg state, the optimizer state, and the training configuration.

Why the decoder is removed. It was trained with a reconstruction objective on raw, non-defaced source images, and the OASIS Data Use Terms carried in the access gate above bar redistributing derivative works or derivative data that could be used to reconstruct a facial image. A ready-to-run reconstruction decoder over face-bearing data is the shape that clause names, so it is not published.

The functional cost is nil, and this is measured rather than assumed. In the published recipe the segmentation finetune runs with load_decoder: False, so it never reads a pretrained decoder tensor. The classification and regression finetunes build a different network whose decoder is a two-tensor linear head (decoder.fc.weight, decoder.fc.bias); its key names do not overlap the pretraining decoder's 52 tensors at all, so those recipes could not consume the pretrained decoder either. Nothing in the published recipe is worse off without it.

Loading.

import torch
ckpt = torch.load("lejepa_fomo300k_encoder.ckpt", map_location="cpu")
state_dict = {k[len("model."):]: v for k, v in ckpt["state_dict"].items()}

Or pass the file straight to the published recipe, unchanged:

CKPT_PATH=/path/to/lejepa_fomo300k_encoder.ckpt ./run_finetune.sh t3

Loading an encoder-only file into the full network leaves the decoder at its random initialization, which is what the recipe expects. Loaders that silently ignore missing keys will not tell you this, so check the key sets yourself if you deviate from the published recipe.

6. What is deliberately absent, and why

• No task-finetuned weights and no inference containers. The challenge's data-usage terms restrict the finetuning data for two of the tasks to use within the competition, so models fitted on it are not ours to redistribute. Only pretraining-stage encoders are here.
• The Methods-track encoder is not in this release (§3). It will be added.
• Further pretraining arms — the objective-weight ablations and the intermediate training trajectory snapshots — are not in this release. They exist and are available on request.

7. Licence

CC BY-NC-SA 4.0, with access gated on accepting the terms above.

This is deliberately a different licence from the code repository, which is Apache-2.0. The weights are derived from FOMO300K, whose CC BY-NC-SA ShareAlike terms carry over to anything built on it, and whose constituent datasets carry Data Use Agreements that must travel with the weights. The code carries no such inheritance. Accepting those terms on this page is what the gate is for.

Cite the LeJEPA paper, the asparagus framework and the FOMO26 challenge as well; the code repository's NOTICE lists what this work builds on.

8. Citation

Chunyi (Eric) Liu, Joshua Shun Liu, Andreia V. Faria.
FOMO26 FariaTeam: LeJEPA pretraining for brain MRI foundation models.
https://github.com/cliu238/fomo26-fariateam
https://doi.org/10.5281/zenodo.22649910

10.5281/zenodo.22649910 is the concept DOI and always resolves to the latest version.