Abstract
Thus far, there have been no reported specific rules for systematically determining the appropriate augmented sample size to optimize model performance when conducting data augmentation. In this paper, we report on the feasibility of synthetic data augmentation using generative adversarial networks (GAN) by proposing an automation pipeline to find the optimal multiple of data augmentation to achieve the best deep learning-based diagnostic performance in a limited dataset. We used Waters' view radiographs for patients diagnosed with chronic sinusitis to demonstrate the method developed herein. We demonstrate that our approach produces significantly better diagnostic performance parameters than models trained using conventional data augmentation. The deep learning method proposed in this study could be implemented to assist radiologists in improving their diagnosis. Researchers and industry workers could overcome the lack of training data by employing our proposed automation pipeline approach in GAN-based synthetic data augmentation. This is anticipated to provide new means to overcome the shortage of graphic data for algorithm training.
Introduction
Deep learning models have revolutionized medical imaging, yet their performance is often throttled by the scarcity of high-quality, labeled datasets. In specialized fields like otorhinolaryngology, collecting large volumes of radiographs is both costly and time-intensive. While Generative Adversarial Networks (GANs) offer a way to synthesize new training data, a critical question remains: how much synthetic data is optimal? This study introduces an automated pipeline to determine the "sweet spot" for data augmentation, specifically applied to chronic sinusitis diagnosis using Waters' view radiographs.
Methodology
The researchers developed a three-stage pipeline to optimize diagnostic performance:
- Image Generation: An Auxiliary Classifier GAN (ACGAN) was employed to generate synthetic images with corresponding class labels (healthy vs. sinusitis), ensuring the synthetic data was categorized correctly from the start.
- Transfer Learning: The team utilized CheXNet, a model pre-trained on over 100,000 chest X-rays, and fine-tuned it for paranasal sinus structures to leverage existing feature recognition capabilities.
- Automated Optimization (Sliding Window): This core innovation iteratively increases the synthetic data volume (multiple $K$) while monitoring the Area Under the Curve (AUC). The process stops when performance gains plateau, identifying the optimal augmentation multiple.
Results and Discussion
The study utilized an internal dataset of 389 image patches and an external dataset of 132 patches for validation. The automated pipeline identified that a 14x multiplier of the original dataset size yielded the peak diagnostic performance.
As shown in the figure above, performance metrics like Accuracy and AUC stabilize after reaching the 14x threshold. Beyond this point, adding more synthetic data leads to diminishing returns, likely due to overfitting or mode collapse within the GAN architecture.
Comparative Performance
The proposed method, termed Optimal Multiple of GAN-based Data Augmentation (OMGDA), significantly outperformed baseline configurations, including conventional geometric augmentation (flipping and rotating).
| Model Configuration | Accuracy | AUC |
|---|---|---|
| Original Data (OD) | 0.785 | 0.855 |
| GAN Class Balancing (GCB) | 0.798 | 0.887 |
| GCB + Conventional Augmentation | 0.810 | 0.908 |
| Proposed Pipeline (GCB + OMGDA) | 0.899 | 0.963 |
The model also demonstrated robust generalization, achieving an AUC of 0.924 on the external test set, confirming its utility across different clinical environments and patient populations.
Conclusion
This research provides a systematic framework for medical AI developers to overcome data shortages. By automating the search for the optimal synthetic data volume, the proposed pipeline maximizes diagnostic accuracy while minimizing the risks of over-augmentation. This approach is expected to be a valuable tool for improving radiologic diagnosis in data-limited scenarios, providing new means to overcome the shortage of graphic data for algorithm training.
Summary updated