FLARE-MedFM/FLARE-Task3-DomainAdaption

FLARE Task3 Domain Adaptation Dataset

Data Description

This is the dataset for MICCAI FLARE 2024-2025 Task3: Unsupervised Domain Adaptation for Abdominal Organ Segmentation in MRI and PET Scans The participants are encouraged to develop efficient abdominal organ segmentation models for MRI and PET scans with labeled and pseudo-labeled CT scans and unlabeled MRI and PET scans.

The training set holds 2050 CT scans, where 50 cases have ground-truth labels from the FLARE22 dataset, and the remaining 2000 cases have pseudo labels generated by the FLARE 2022 winning solution. For domain adaptation, the training set also contains 4817 unlabeled MRI scans and 1000 unlabeled PET scans. For those participants who are constrained by computing resources, we also provide an unlabeled core set to develop the methods, where 100 unlabeled MRI and 100 unlabeled PET scans are sampled from the original MRI and PET training set.

The validation set contains 110 MRI scans and 50 PET scans, while the testing set contains 300 MRI scans and 200 PET scans.

Task Setting

The participants should develop two domain adaptive models, one for MRI segmentation and the other for PET segmentation.

In MRI scans, there are 13 organs and tissues for segmentation, including the liver (labeled 1), right kidney (labeled 2), spleen (labeled 3), pancreas (labeled 4), aorta (labeled 5), inferior vena cava (IVC, labeled 6), right adrenal gland (RAG, labeled 7), left adrenal gland (LAG, labeled 8), gallbladder (labeled 9), esophagus (labeled 10), stomach (labeled 11), duodenum (labeled 12), left kidney (labeled 13).

In PET scans, there are 4 organs for segmentation, including the liver (labeled 1), right kidney (labeled 2), spleen (labeled 3), left kidney (labeled 4).

Data Structure

train_CT_gt_label： 50 CT scans with ground-truth labels.

train_CT_pseudolabel： 2000 CT scans with pseudo labels generated by the FLARE 2022 winning teams, i.e., aladdin5 and blackbean.

train_MRI_unlabeled： 4817 unlabeled MRI scans

train_PET_unlabeled： 1000 unlabeled PET scans

coreset_train_unlabeled_MRI_PET： 100 unlabeled MRI and 100 unlabeled PET scans sampled from the original unlabeled MRI and PET training set.

validation： 160 MRI scans and 50 PET scans

FLARE-Task3-DomainAdaption/
├── coreset_train_unlabeled_MRI_PET/
│        ├── MRI_unlabeled_100_random/
│        └── PET_unlabeled_100_random/
├── train_CT_gt_label/
│        ├── imagesTr/
│        ├── labelsTr.7z
│        └── dataset.json
├── train_CT_pseudolabel/
│        ├── imagesTr/
│        ├── pseudo_label_aladdin5_flare22.7z
│        └── pseudo_label_blackbean_flare22.zip
├── train_MRI_unlabeled/
│        ├── AMOS-833/
│        └── LLD-MMRI-3984/
├── train_PET_unlabeled/
├── validation/
│        ├── MRI_imagesVal/
│        ├── MRI_labelsVal/
│        ├── PET_imagesVal/
│        ├── PET_labelsVal/
│        └── readme.txt
└── README.md

Data Download

pip install -U huggingface_hub
huggingface-cli download FLARE-MedFM/FLARE-Task3-DomainAdaption --repo-type dataset --local-dir ./FLARE-MedFM/FLARE-Task3-DomainAdaption --local-dir-use-symlinks False