Related Experiment Video
Updated: Sep 24, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.0K
SinGAN-Seg: Synthetic training data generation for medical image segmentation
Vajira Thambawita1,2, Pegah Salehi1, Sajad Amouei Sheshkal1
1SimulaMet, Oslo, Norway.
Plos One
|May 2, 2022
Summary
This study introduces SinGAN-Seg, a novel pipeline for generating synthetic medical images and masks from a single training image. SinGAN-Seg effectively enhances medical image segmentation model performance, especially with limited real data.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Medical data analysis for abnormalities is time-consuming and expensive, especially for rare conditions.
- Machine learning models require large datasets, which are difficult to obtain in medicine due to privacy, annotation costs, and data scarcity.
- Existing generative adversarial networks (GANs) often need extensive training data.
Purpose of the Study:
- To present SinGAN-Seg, a novel synthetic data generation pipeline for medical images and segmentation masks.
- To demonstrate the pipeline's ability to create artificial datasets when real data sharing is restricted.
- To evaluate the quality and utility of synthetic data for training medical image segmentation models.
Main Methods:
- Developed SinGAN-Seg, a generative adversarial network (GAN) model trained on a single medical image and its corresponding ground truth mask.
- Employed a style transfer technique to enhance the quality of generated synthetic medical images.
- Compared SinGAN-Seg against state-of-the-art GANs for synthetic image preparation with limited training data.
Main Results:
- SinGAN-Seg successfully generated high-quality synthetic medical images with accurate segmentation masks using only one training sample.
- Qualitative and quantitative evaluations showed improved data quality compared to other GANs, particularly with limited datasets.
- Models trained with SinGAN-Seg synthetic data achieved performance comparable to models trained on real data, and showed significant improvement when real data was scarce.
Conclusions:
- SinGAN-Seg offers an effective solution for generating synthetic medical data, addressing challenges of data scarcity and privacy.
- The pipeline enhances the performance of medical image segmentation models, especially in low-data regimes.
- The publicly available code and open dataset facilitate further research and application in medical AI.
