procodec/cvpr26-vo-eao

CVPR 2026 CT-FM Challenge — Task 2 EAO, Vision-Only Track Submission

Submission artifact for Aviral (team leader: Aviral Kaintura) to the CVPR 2026: Foundation Models for General CT Image Diagnosis challenge (Codabench competition 12650), Task 2 — Embedding Aggregation Optimization (EAO), Vision-Only (VO) track.

Validation result: 0.7086 balanced accuracy / 0.728 AUROC, mean over 17 binary classification targets (Codabench submission #715408).

Contents

• aviral_vo_eao.tar.gz — Docker image (aviral_vo_eao:latest, ~13GB). Wraps the organizer-provided CT-NEXUS backbone as a frozen feature extractor; entrypoint extract_feat_EAO.sh produces per-case spatial embeddings.
• aviral_vo_eao_inference.zip — training script (run_EAO_improved.py), ensemble inference script (predict_eao_ensemble.py), trained checkpoints (4 seeds × 17 targets × multiple learning rates), and a standalone threshold-optimization script.
• predict_eao_ensemble.py — standalone copy of the final inference script, with per-target decision thresholds merged in directly so it runs as a single step (no separate threshold-optimization pass required).

License

The team's own contributions in this repository — training code, inference/ensembling code, threshold-calibration logic, and trained checkpoint weights — are released under the MIT License.

This does not cover the CT-NEXUS backbone bundled inside the Docker image. CT-NEXUS is an organizer-provided baseline for this challenge (github.com/kmin940/CT-NEXUS) with no separate publication, model card, or license file of its own. It is included here only as required to reproduce this submission's Docker container; if you intend to reuse the backbone itself outside the context of this challenge, refer to its source repository directly rather than assuming MIT terms apply to it.

Citation

If you reference this submission, please cite the challenge itself:

CVPR 2026 Workshop — Foundation Models for General CT Image Diagnosis, Codabench competition 12650.