Neck-Ultrasound-Dataset

NNUS: Dataset for Normal Neck Organ Detection in Ultrasound

Summary statement

Ultrasound images of the neck are often noisy and artifacts, requiring extensive experience to accurately identify the human organ. Therefore, image detection models trained on a large dataset for Normal Neck Ultrasound (NNUS) images have the potential to help radiologists and sonographers in real time. It can also be used for educational purposes for students learning radiology and can be usefully used to distinguish it from normal people by adding a dataset containing abnormal neck lesions.

The anatomical appearance of the normal person's neck and ultrasound images of the organs and its label dataset can shorten the time for many researchers to scan and collect the normal person's neck for their research. In addition, many researchers can develop new models using their datasets and extra-validate their studies using these datasets.

Abstract

Our NNUS dataset consists of 790 normal human neck ultrasound images and their labels.
We trained 100 epochs using YOLOv5s pretrained model and obtained mAP 0.8044, precision 0.7952, recall 0.8082.
For the validation of the model we have learned, we input the newly scanned normal person's neck ultrasound video, and we were able to obtain the video which detected organs in the neck area as shown below.

[Click on the image below to play the video]

Structure of Dataset
Usage
Results of Training
Acknowledgements
Citation

Structure of Dataset

    NNUS/
    ├── images
    │   ├── 0001.jpg       <- Ultrasound images
    │   ├── 0002.jpg
    │   └── 0003.jpg
    └── labels
        ├── 0001.txt       <- Label 
        ├── 0002.txt
        ├── 0003.txt
        └── classes.txt    <- Class

Images

All images are composed of 494×447 pixel jpg format files. All shots were taken from 25 volunteers (average age 24.5 ± 3.5 years), presumed to be disease-free, ultrasound images of the neck.

One radiologic technologist (1 year of experience) and three sonography students in Shingu College (Seongnam, Republic of Korea) participated in the scan. This dataset is the result of scanning and extracting images using LOGIQ P6 (GE healthcare, Connecticut, U.S.). Three types of scanning methods (longitudinal, transverse, and oblique scan) were randomly applied, and scanned about 30 to 40 images per person.

Labels

Two radiologic technologists (11 years and 6 years of experience) used Label-Img (https://github.com/tzutalin/labelImg) to label the trachea, thyroid, common carotid artery, internal jugular vein, esophagus, longus colli muscle, strap muscle, and sternocleidomastoid muscle with bounding boxes.

The location information of these boxes was converted into coordinates and saved as a text file. The total number of labels is as follows.

Usage

AI using datasets can be trained using object detection models such as YOLO.

We identified training performance using the YOLOv5s pre-learning model (YOLOv5 by Ultralytics (Version 7.0)).

Open the nnus_train_and_valid.ipynb file on Github and click the "Open in Colab" button.
Execute the code line by line from the top. (Press the left play button on each code, or click on the code and press Shift+Enter.)
When you run python train.py, you can change the parameters you want and proceed with training. (For example, you can change the weight from yolo5s.pt to yolo5m.pt.)
When train.py is all done, you can get a model file named /content/yolov5/runs/detect/exp/weight/best.pt.
Run detect.py. Test.mp4, which was never used for training, is detected through the best.pt model.
After detect.py is completed, open /content/yolov5/runs/val/exp/test.mp4 to visually validate whether human organs are well detected.

Results of Training

best epoch: 90
mAP_0.5: 0.8044
mAP_0.5:0.95: 0.4717
precision: 0.7952
recall: 0.8082

Acknowledgements

This code borrows from

Jocher, G. (2020). YOLOv5 by Ultralytics (Version 7.0) [Computer software]. https://doi.org/10.5281/zenodo.3908559
Tzutalin. LabelImg. Git code (2015). https://github.com/tzutalin/labelImg

License

CC BY-NC-SA 4.0

When publishing a research paper using a dataset, the following citations must be included.

Citation

Joo Young Oh, Euihwan Jeong, Moo Jin Jeong, Joo Young Lee, Hoon-Hee Park (2023). NNUS: Large Dataset for Normal Neck Organ Detection in Ultrasound. https://github.com/joo053/Neck-Ultrasound-Dataset

joo053/Neck-Ultrasound-Dataset