Related Experiment Videos
Generative adversarial networks for balancing and expanding data resources for computer-aided detection in
J J H van der Laan1,2, J van Lune3,4, L R B Schomaker3
1Department of Gastroenterology and Hepatology, University Medical Center Groningen, University of Groningen, Groningen, Netherlands.
Frontiers in Medical Technology
|July 24, 2026
Summary
A modified StyleGAN2-ADA can generate synthetic colonoscopy images, improving training data for computer-aided detection (CADe) algorithms. This approach creates balanced datasets, enhancing CADe performance comparable to real-world data.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer-Aided Detection
Background:
- High-quality colonoscopy data is crucial for developing computer-aided detection (CADe) algorithms.
- Real-world datasets are costly and often lack sufficient representation of flat adenomas, leading to imbalanced data.
- Generative adversarial networks (GANs) show potential for creating synthetic medical images.
Purpose of the Study:
- To evaluate StyleGAN2-ADA's ability to generate synthetic colonoscopy images for CADe training.
- To assess if a modified StyleGAN2-ADA can improve synthesis control for balanced datasets.
- To compare the performance of CADe models trained on synthetic versus real-world data.
Main Methods:
- Two synthetic datasets were generated using original and modified StyleGAN2-ADA with feature-clustered conditioning vectors.
- Generative adversarial networks were trained on local colonoscopy images and evaluated using Fréchet Inception Distance.
- Three CADe models were trained on synthetic or real data and tested on external polyp image datasets, measuring performance with mean average precision (mAP).
Main Results:
- The modified StyleGAN2-ADA achieved a significantly lower Fréchet Inception Distance (7.54) compared to the original (13.68).
- CADe models trained on synthetic data from the modified StyleGAN2-ADA demonstrated performance comparable to models trained on real-world data (p > 0.05).
- The modified synthetic data approach outperformed the original StyleGAN2-ADA synthetic data, yielding higher mAP scores (0.77 ± 0.03 vs. 0.64 ± 0.02 and 0.91 ± 0.02 vs. 0.87 ± 0.02).
Conclusions:
- Modified StyleGAN2-ADA with feature-clustered conditioning vectors effectively synthesizes balanced colonoscopy datasets.
- These synthetic datasets serve as valuable alternative training resources for CADe algorithms.
- This approach enhances the development and validation of CADe systems, addressing data imbalance issues.