Docker container (with all model weights baked in) for our submission to the MICCAI 2026 REG Challenge (pathology report generation from whole-slide images).

This repository exists to satisfy the organizers' requirement for a publicly accessible link to the Docker container / model weights. The artifact here is byte-identical to the image evaluated on the challenge platform.

The submission consists of two Grand Challenge artifacts: the algorithm container image and the model tarball mounted at /opt/ml/model. Both are documented below.

All weights trained by us — the per-organ VIVI recognition heads — are inside this image.

The model.tar.gz attached to our Grand Challenge algorithm contains no weights of our own. It is nothing more than a HuggingFace cache directory holding one publicly available third-party checkpoint:

We deliberately do not re-host these weights. CONCH is distributed under CC-BY-NC-ND-4.0 behind a gated access agreement, and mirroring it here would bypass the authors' access controls. Reviewers can obtain the identical checkpoint directly from MahmoodLab after accepting their terms, and rebuild the tarball:

Pinning the revision guarantees the same weights we ran with. (The gzip archive itself will not be bit-identical, since tar records file timestamps.)

The image follows the challenge's standard I/O contract: it mounts an input directory containing the whole-slide image and writes the generated report to the output directory. It runs fully offline — no network access is required or attempted at inference time.

Nothing is downloaded at runtime; every weight is either baked into the image or supplied by the mounted model tarball described above. The image itself carries a publicly available vision-language model used by the Metric-B stage, plus our own per-organ recognition heads, which were trained on the challenge training split only. Exact model identities, versions and load paths are recorded in the source repository linked below.

VIVI is a flat attribute-library recognizer over CONCH-v1 coverage embeddings. One dispatcher routes each case by organ to a per-organ recognizer, whose predicted attribute fields are then mapped through a deterministic, data-mined canonical chain-of-thought DAG to produce the final report. The DAG structure is derived from training-set report templates; only the recognition layer is learned.

Seven organs are covered (lung, colon, stomach, bladder, prostate, breast, cervix) across 121 diagnoses and 158 recognition nodes.

Built and evaluated for the challenge's GPU targets (NVIDIA T4 / A10G), CUDA 12.1, PyTorch 2.4.1.

Training and evaluation code is in a separate repository, shared privately with the challenge organizers as required by the submission rules: https://github.com/cloud1710/VIVI_CoT

Released under CC BY 4.0, per the REG2026 challenge submission requirements. Bundled third-party model weights remain under their own respective licenses.