Related Experiment Video
Updated: Oct 29, 2025
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
589
Effects of sample size and data augmentation on U-Net-based automatic segmentation of various organs
Takafumi Nemoto1, Natsumi Futakami2, Etsuo Kunieda3,2
1Department of Radiology, Keio University School of Medicine, Shinanomachi 35, Shinjuku-ku, Tokyo, 160-8582, Japan. takatohoku@gmail.com.
Radiological Physics and Technology
|July 13, 2021
Summary
Deep learning models like U-Net improve radiation therapy planning through automatic segmentation. Data augmentation, particularly horizontal flipping, significantly boosts performance, especially when limited computed tomography imaging data is available.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Radiation Oncology
Background:
- Automatic segmentation is vital for radiation therapy planning.
- Medical imaging data collection and labeling present significant challenges.
- Deep learning methods, such as U-Net, show promise for segmentation tasks.
Purpose of the Study:
- To investigate the impact of sample size on U-Net segmentation accuracy.
- To evaluate the effectiveness of data augmentation techniques.
- To determine optimal dataset sizes for reliable segmentation in radiation therapy planning.
Main Methods:
- U-Net deep learning model was employed for segmentation.
- Computed tomography images from chest and pelvic regions were analyzed.
- Sample sizes ranged from 10 to 200 (chest) and 10 to 500 (pelvic) cases.
- Horizontal-flip data augmentation was compared against no augmentation.
Main Results:
- Dice Similarity Coefficient (DSC) improved with increased sample size, stabilizing after approximately 200 cases.
- Horizontal-flip augmentation was nearly as effective as doubling the dataset size.
- Data augmentation helped stabilize DSCs at smaller sample sizes for most organs.
Conclusions:
- Data augmentation is crucial for improving deep learning segmentation with limited medical imaging datasets.
- Findings are particularly relevant for automating radiation therapy in rare cancers.
- Optimizing sample size and employing augmentation can enhance segmentation reliability.
