Modelpublic

hieuphamha/A-nnU-Net-based-asymmetric-supervision-strategy

sourceHugging Faceapache-2.0updated 24d agoView on Hugging Face

0likes

Model Card

OS50 + AURA model weights

This model is presented in the paper Asymmetric Paired-Annotation Learning for Multi-Structure ULF Pediatric Brain MRI Segmentation.

Pretrained checkpoints for AURA — Asymmetric Paired-Annotation Learning for ULF Pediatric Brain MRI, the LISA 2026 Task 2 challenge report.

The release contains the two fold-0 3d_fullres branches used by the packaged submission:

Branch	Dataset	Trainer	Fusion weight
OS50	`Dataset001_LISA`	`nnUNetTrainerDiceTopK10_OS50_LS005`	0.6
AURA	`Dataset002_LISA_VCF`	`nnUNetTrainerAURA_v0`	0.4

Download the complete bundle:

bash

hf download hieuphamha/A-nnU-Net-based-asymmetric-supervision-strategy \
  --local-dir aura-models

Inference

The released system is a two-model ensemble and is intended to be run through the submission Docker container. It requires a Linux host with an NVIDIA GPU, Docker, and the NVIDIA Container Toolkit. Only the ULF MRI is required at inference; no HF image or segmentation mask is used.

bash

# 1. Get the inference code.
git clone https://github.com/minhdang050806/A-nnU-Net-based-asymmetric-supervision-strategy.git
cd A-nnU-Net-based-asymmetric-supervision-strategy

# 2. Download both checkpoints and their nnU-Net metadata.
hf download hieuphamha/A-nnU-Net-based-asymmetric-supervision-strategy \
  --local-dir aura-models
mkdir -p docker/models
cp -a aura-models/models/. docker/models/

# 3. Verify every packaged model file.
(cd docker && sha256sum -c SHA256SUMS)

# 4. Add nnU-Net v2 to the Docker build context and install AURA extensions.
git clone https://github.com/MIC-DKFZ/nnUNet.git docker/nnUNet
python scripts/install_extensions.py --nnunet-dir docker/nnUNet
python scripts/install_extensions.py --nnunet-dir docker/nnUNet --check

# 5. Build the image.
cd docker
./build.sh lisa-os50-aura:local

# 6. Run inference. Use absolute input and output paths.
./test.sh lisa-os50-aura:local /absolute/path/to/input /absolute/path/to/output

The input directory is searched recursively for NIfTI volumes named <case>_ciso.nii.gz, <case>_0000.nii.gz, or <case>.nii.gz. The container writes one mask per case as <case>_seg_prediction.nii.gz. Output labels are 0 (background) and 1–11 (foreground structures), and the input image geometry is preserved.

Internally, both fold-0 3d_fullres models predict soft probabilities from the ULF volume. The final mask is argmax(0.6 * p_OS50 + 0.4 * p_AURA), followed by retaining the largest 26-connected component independently for labels 1–11.

For additional runtime details and troubleshooting, see `docker/README.md`.

Checksums

text

422d2b25b0d88712896c7c58057d4e0368ca05b0a9d7557b80e405d331a427e0  models/Dataset001_LISA/nnUNetTrainerDiceTopK10_OS50_LS005__nnUNetPlans__3d_fullres/fold_0/checkpoint_best.pth
701674f232443434bd48ce921011c5f69decf85ffc4b2a0d6d54654f721dfb2c  models/Dataset002_LISA_VCF/nnUNetTrainerAURA_v0__nnUNetPlans__3d_fullres/fold_0/checkpoint_best.pth

The AURA file is the surviving epoch-458 packaged checkpoint (cross-evaluation DSC 0.796101), not the overwritten intermediate snapshot reported as the standalone AURA result in the paper.