We are proud to introduce AbdomenAtlas-8K, a substantial multi-organ dataset with the spleen, liver, kidneys, stomach, gallbladder, pancreas, aorta, and IVC annotated in 8,448 CT volumes, totaling 3.2 million CT slices.
An endeavor of such magnitude would demand a staggering 1,600 weeks or roughly 30.8 years of an experienced annotator's time.
In contrast, our annotation method has accomplished this task in three weeks (premised on an 8-hour workday, five days a week) while maintaining a similar or even better annotation quality.
AbdomenAtlas-8K: Annotating 8,000 CT Volumes for Multi-Organ Segmentation in Three Weeks
Chongyu Qu1, Tiezheng Zhang1, Hualin Qiao2, Jie Liu3, Yucheng Tang4, Alan L. Yuille1, and Zongwei Zhou1,*
1 Johns Hopkins University,
2 Rutgers University,
3 City University of Hong Kong,
4 NVIDIA
NeurIPS 2023
paper | code | dataset | annotation
AbdomenAtlas-8K: Human-in-the-Loop Annotating Eight Anatomical Structures for 8,448 Three-Dimensional Computed Tomography Volumes in Three Weeks
Chongyu Qu1, Tiezheng Zhang1, Hualin Qiao2, Jie Liu3, Yucheng Tang4, Alan L. Yuille1, and Zongwei Zhou1,*
1 Johns Hopkins University,
2 Rutgers University,
3 City University of Hong Kong,
4 NVIDIA
RSNA 2023 (Oral Presentation)
paper | code | slides
★ We have maintained a document for Frequently Asked Questions.
git clone https://github.com/MrGiovanni/AbdomenAtlas
See installation instructions to create an environment and obtain requirements.
We offer pre-trained checkpoints of Swin UNETR and U-Net. The models were trained on a combination of 14 publicly available CT datasets, consisting of 3,410 (see details in CLIP-Driven Universal Model).
Download the trained models and save them into ./pretrained_checkpoints/
.
Architecture | Param | Download |
---|---|---|
U-Net | 19.08M | link |
Swin UNETR | 62.19M | link |
It can be publicly available datasets (e.g., BTCV) or your private datasets. Currently, we only take data formatted in nii.gz
. This repository will help you assign annotations to these datasets, including 25 organs and six types of tumors (where the annotation of eight organs is pretty accurate).
Taking the BTCV dataset as an example, download this dataset and save it to the datapath
directory.
cd $datapath
wget https://www.dropbox.com/s/jnv74utwh99ikus/01_Multi-Atlas_Labeling.tar.gz
tar -xzvf 01_Multi-Atlas_Labeling.tar.gz
Generate a list for this dataset.
cd AbdomenAtlas/
python -W ignore generate_datalist.py --data_path $datapath --dataset_name $dataname --folder img --out ./dataset/dataset_list --save_file $dataname.txt
CUDA_VISIBLE_DEVICES=0 python -W ignore test.py --resume pretrained_checkpoints/unet.pth --backbone unet --save_dir $savepath --dataset_list $dataname --data_root_path $datapath --store_result >> logs/$dataname.unet.txt
CUDA_VISIBLE_DEVICES=0 python -W ignore test.py --resume pretrained_checkpoints/swinunetr.pth --backbone swinunetr --save_dir $savepath --dataset_list $dataname --data_root_path $datapath --store_result >> logs/$dataname.swinunetr.txt
To generate attention maps for the active learning process (Step 5 [optional]), remember to save entropy and soft predictions by using the options --store_entropy
and --store_soft_pred
In the assembly process, our utmost priority is given to the original annotations supplied by each public dataset. Subsequently, we assign secondary priority to the revised labels from our annotators. The pseudo labels, generated by AI models, are accorded the lowest priority. The following code can implement this priority into the assembled dataset.
python -W ignore assemble.py --data_path $savepath --dataset_name $dataname --backbone swinunetr --save_dir SAVE_DIR --version V1
This is how our AbdonmenAtlas-8K appears
$savepath/
├── $dataname_img0001
├── $dataname_img0002
├── $dataname_img0003
│── ct.nii.gz
├── original_label.nii.gz
├── pseudo_label.nii.gz
└── segmentations
├── spleen.nii.gz
├── liver.nii.gz
├── pancreas.nii.gz
If you want to perform the active learning process, you will need the following active learning instructions to generate the attention map for human annotators.
Figure. Illustration of an attention map.- Release pre-trained AI model checkpoints (U-Net and Swin UNETR)
- Release the AbdomenAtlas-8K dataset (we commit to releasing 3,410 of the 8,448 CT volumes)
- Support more data formats (e.g., dicom)
@article{qu2023annotating,
title={Annotating 8,000 Abdominal CT Volumes for Multi-Organ Segmentation in Three Weeks},
author={Qu, Chongyu and Zhang, Tiezheng and Qiao, Hualin and Liu, Jie and Tang, Yucheng and Yuille, Alan and Zhou, Zongwei},
journal={arXiv preprint arXiv:2305.09666},
year={2023}
}
This work was supported by the Lustgarten Foundation for Pancreatic Cancer Research and partially by the Patrick J. McGovern Foundation Award. We appreciate the effort of the MONAI Team to provide open-source code for the community.