UNGT Dataset & AAMS Model

Official PyTorch Implementation for Knowledge-Based Systems Submission “UNGT: Ultrasound Nasogastric Tube Dataset for Medical Image Analysis”

Dataset

The UNGT dataset includes 493 images gathered from 110 patients with an average image resolution of approximately 879 $\times$ 583. To the best of our knowledge, this is the first publicly available medical ultrasound dataset related to the nasogastric tube

Researchers are free to download the dataset from ./data/ungt/images and ./data/ungt/masks for academic purposes. The masks comprise grey levels from 1 to 4, corresponding to the liver, stomach, tube, and pancreas

Model

The AAMS model addresses data limitation and imbalance concurrently

It outperforms existing methods by varying degrees, particularly on infrequent or minor structures

To deploy our AAMS model in your own scenarios, please install the core dependencies using

pip install -r requirements.txt

With the core and required additional dependencies installed, please ensure that your project structure follows

├── AAMS
    ├── data
    |   ├── ungt
    |   |   ├── images
    |   |   ├── masks
    |   |   ├── train.txt
    |   |   └── val.txt
    |   ├── camus
    |   └── ...
    ├── src
    ├── checkpoint
    └── ...

We have provided our training and validation lists in ./data/ungt/train.txt and ./data/ungt/val.txt for reproducibility. To generate new splits, execute

python split.py

The training can be performed with

python train.py

Citation

If you use the UNGT dataset or the AAMS model in your research, do remember to cite our paper as

@article{liu2025ungt,
  title={UNGT: Ultrasound Nasogastric Tube Dataset for Medical Image Analysis},
  author={Liu, Zhaoshan and Lee, Chau Hung and Lv, Qiujie and Wee, Nicole Kessa and Shen, Lei},
  journal={Knowledge-Based Systems},
  volume={330},
  pages={114615},
  year={2025},
  doi={10.1016/j.knosys.2025.114615}
}