UltraSam: A Foundation Model for Ultrasound using Large Open-Access Segmentation Datasets

Adrien Meyer, Aditya Murali, Didier Mutter, Nicolas Padoy

Usage

Click to expand Install

You may need to install a specific version of PyTorch, depending on your hardware.

Create a conda environment and activate it.

conda create --name UltraSam python=3.8 -y
conda activate UltraSam

Install the OpenMMLab suite and other dependencies

pip install -U openmim
mim install mmengine
mim install "mmcv>=2.0.0"
mim install mmdet
mim install mmpretrain

If you wish to process the datasets;

pip install SimpleITK
pip install scikit-image
pip install scipy

Pre-trained UltraSam model checkpoint is accessible at this link.

To train / test, you will need a coco.json annotation file, and create a symbolik link to it, or modify the config files to point to your annotation file.

To train from scratch, you can use the code in weights to download and convert SAM, MEDSAM and adapters weights.

In local, inside UltraSam repo;

export PYTHONPATH=$PYTHONPATH:.

mim train mmdet configs/UltraSAM/UltraSAM_full/UltraSAM_point_refine.py --gpus 4 --launcher pytorch --work-dir ./work_dirs/UltraSam
mim test mmdet configs/UltraSAM/UltraSAM_full/UltraSAM_point_refine.py --checkpoint ./work_dirs/UltraSam/iter_30000.pth
mim test mmdet configs/UltraSAM/UltraSAM_full/UltraSAM_box_refine.py --checkpoint ./work_dirs/UltraSam/iter_30000.pth


mim train mmpretrain configs/UltraSAM/UltraSAM_full/downstream/classification/BUSBRA/resnet50.py \
    --work-dir ./work_dirs/classification/BUSBRA/resnet
mim train mmpretrain configs/UltraSAM/UltraSAM_full/downstream/classification/BUSBRA/MedSAM.py \
    --work-dir ./work_dirs/classification/BUSBRA/MedSam
mim train mmpretrain configs/UltraSAM/UltraSAM_full/downstream/classification/BUSBRA/SAM.py \
    --work-dir ./work_dirs/classification/BUSBRA/Sam
mim train mmpretrain configs/UltraSAM/UltraSAM_full/downstream/classification/BUSBRA/UltraSam.py \
    --work-dir ./work_dirs/classification/BUSBRA/UltraSam
mim train mmpretrain configs/UltraSAM/UltraSAM_full/downstream/classification/BUSBRA/ViT.py \
    --work-dir ./work_dirs/classification/BUSBRA/ViT

mim train mmdet configs/UltraSAM/UltraSAM_full/downstream/segmentation/BUSBRA/resnet.py \
    --work-dir ./work_dirs/segmentation/BUSBRA/resnet
mim train mmdet configs/UltraSAM/UltraSAM_full/downstream/segmentation/BUSBRA/UltraSam.py \
    --work-dir ./work_dirs/segmentation/BUSBRA/UltraSam_3000
mim train mmdet configs/UltraSAM/UltraSAM_full/downstream/segmentation/BUSBRA/SAM.py \
    --work-dir ./work_dirs/segmentation/BUSBRA/SAM
mim train mmdet configs/UltraSAM/UltraSAM_full/downstream/segmentation/BUSBRA/MedSAM.py \
    --work-dir ./work_dirs/segmentation/BUSBRA/MedSAM

US-43d

Ultrasound imaging presents a substantial domain gap compared to other medical imaging modalities; building an ultrasound-specific foundation model therefore requires a specialized large-scale dataset. To build such a dataset, we crawled a multitude of platforms for ultrasound data. We arrived at US-43d, a collection of 43 datasets covering 20 different clinical applications, containing over 280,000 annotated segmentation masks from both 2D and 3D scans.

Click to expand datasets table

Dataset	Link
105US	researchgate
AbdomenUS	kaggle
ACOUSLIC	grand-challenge
ASUS	onedrive
AUL	zenodo
brachial plexus	github
BrEaST	cancer imaging archive
BUID	qamebi
BUS_UC	mendeley
BUS_UCML	mendeley
BUS-BRA	github
BUS (Dataset B)	mmu
BUSI	HomePage
CAMUS	insa-lyon
CardiacUDC	kaggle
CCAUI	mendeley
DDTI	github
EchoCP	kaggle
EchoNet-Dynamic	github
EchoNet-Pediatric	github
FALLMUD	kalisteo
FASS	mendeley
Fast-U-Net	github
FH-PS-AOP	zenodo
GIST514-DB	github
HC	grand-challenge
kidneyUS	github
LUSS_phantom	Leeds
MicroSeg	zenodo
MMOTU-2D	github
MMOTU-3D	github
MUP	zenodo
regPro	HomePage
S1	ncbi
Segthy	TUM
STMUS_NDA	mendeley
STU-Hospital	github
TG3K	github
Thyroid US Cineclip	standford
TN3K	github
TNSCUI	grand-challenge
UPBD	HomePage
US nerve Segmentation	kaggle

Once you downloaded the datasets: Run each converter in datasets/datasets

# run coco converters

# then preprocessing
python datasets/tools/merge_subdir_coco.py
python datasets/tools/split_coco.py
python datasets/tools/create_agnostic_coco.py path_to_datas_root --mode train
python datasets/tools/create_agnostic_coco.py path_to_datas_root --mode val
python datasets/tools/create_agnostic_coco.py path_to_datas_root --mode test
python datasets/tools/merge_agnostic_coco.py path_to_datas_root path_to_datas_root/train.agnostic.noSmall.coco.json --mode train
python datasets/tools/merge_agnostic_coco.py path_to_datas_root path_to_datas_root/val.agnostic.noSmall.coco.json --mode val
python datasets/tools/merge_agnostic_coco.py path_to_datas_root path_to_datas_root/test.agnostic.noSmall.coco.json --mode test

References

If you find our work helpful for your research, please consider citing us using the following BibTeX entry:

@article{meyer2024ultrasam,
  title={UltraSam: A Foundation Model for Ultrasound using Large Open-Access Segmentation Datasets},
  author={Meyer, Adrien and Murali, Aditya and Mutter, Didier and Padoy, Nicolas},
  journal={arXiv preprint arXiv:2411.16222},
  year={2024}
}

---