MVAA 2026 — Team Jmees final-system weights
Trained checkpoints of our solution to the MICCAI 2026 Mitral Valve Anatomy Analysis (MVAA) challenge (final v026 system; hidden-test DSC 0.838 / 0.887 / 0.836 for CT / TEE / video).
- Code: https://github.com/ShunsukeKikuchi/MVAA26-jmees-sol
- Paper: Solution for MVAA challenge 2026: Ensemble, Tool-paste Augmentation and Tool-free View Generation (MWM2026)
Layout
| Folder | Contents | Networks |
|---|---|---|
t1/ |
Task 1 CT members | plainconv+pseudo (a15), SwinUNETR scratch/BTCV (swin_a09/a10), 2.5D yz / oblique / distortion (a12*), 2D-encoder+3D-decoder hybrid (a13) — 5 folds each |
t2/ |
Task 2 TEE members | plainconv (a05) ×5, plainconv+MVSeg-external (a07) ×5 |
t3/ |
Task 3 video members | UNet++/EfficientNet-B7 (a04), MaxViT-L (a09), ConvNeXt-L DINOv3 (a11) — 5 folds each |
nnunet_results/ |
nnU-Net v2 results trees | Dataset001_MitralCT + Dataset002_TEE, 3d_fullres, 5 folds each |
WEIGHTS_MANIFEST.sha256 lists the SHA-256 of every checkpoint:
sha256sum -c WEIGHTS_MANIFEST.sha256
Download
pip install -U huggingface_hub
hf download negichi/MVAA26-jmees-weights --local-dir weights/
The resulting weights/ tree is exactly what submission/v026_final_docker/bundle_weights.sh
in the code repository expects.
License / intended use
Code and weights are released under MIT. The models were trained on MVAA 2026 challenge data (plus the external corpora documented in the paper) for research purposes; they are not medical devices and must not be used for clinical decision-making.