R3GAN-based Optimal Strategy for Augmenting Small Medical Dataset

Tsung-Wei Pan, Chang-Hong Wu, Jung-Hua Wang, Ming-Jer Chen, Yu-Chiao Yi, Tsung-Hsien LeearXiv 2025 · DOI · arXiv

Abstract

Deep learning in medical imaging often faces the "small data" bottleneck, where privacy concerns and high annotation costs limit dataset size. This paper presents a specialized training strategy for the Relativistic Pairing GAN (R3GAN) to generate high-quality synthetic images from limited clinical data. By implementing an extended burn-in phase and a reversed regularization schedule—increasing gradient penalties over time—the researchers achieved an 86% reduction in Fréchet Inception Distance (FID) on a human embryo dataset. These synthetic samples significantly improved the classification of rare developmental stages, increasing the F1-score from 0.11 to 0.60.

Introduction

Medical imaging datasets are frequently characterized by small sample sizes and significant class imbalances. In the context of Time-Lapse Imaging (TLI) for human embryos, certain developmental stages, such as the three-cell (t3) stage, are transient and rare. Traditional data augmentation techniques like rotation or flipping are often ineffective for biological objects with high rotational symmetry. Generative Adversarial Networks (GANs) offer a potential solution by synthesizing realistic new samples; however, standard GAN training often fails on small datasets due to rapid discriminator overfitting.

Methods: Optimized Training Strategy

The authors utilized the R3GAN architecture, which incorporates relativistic pairing and zero-centered gradient penalties ($R1$ and $R2$). To adapt this model for small-scale embryo datasets, three primary modifications were introduced:

1. Extended Burn-in and Delayed Decay

Unlike standard GANs that use a brief warm-up period, this approach employs a burn-in phase lasting 100% to 150% of the typical training duration. This prevents premature model collapse by allowing the generator to capture the basic data distribution before learning fine details. Furthermore, a "Squared Delayed Decay" schedule was implemented to maintain a high learning state for a longer duration.

Comparison of different decay schedules, highlighting the 150% burn-in and delayed decay strategies.

2. Reversed Regularization ($\gamma$) Schedule

In typical large-scale GAN training, the regularization parameter $\gamma$ starts high and decreases. For small datasets, the authors reversed this logic: starting with a low $\gamma$ (e.g., 5 or 15) to encourage exploration, and increasing it (e.g., to 40 or 150) as training progresses. This "handicaps" the discriminator just as it begins to memorize the small training set, maintaining the adversarial balance.

Impact of different Gamma schedules on training progress and FID scores. The reversed schedule (Exp013, 15→150) demonstrates superior stability and lower FID in early-to-mid training.

Experimental Results

The study utilized a dataset of ~3.6k human embryo images from Taichung Veterans General Hospital, focusing on the t2, t3, and t4 stages. The optimized configuration (Exp017) led to a dramatic drop in FID from a baseline of 259 to 36.

When these synthetic images were used to balance the training set for a BLIP-based classifier, the performance on the minority t3 class improved significantly:

  • Recall: Increased from 0.06 to 0.69.
  • F1-Score: Increased from 0.11 to 0.60.

Conclusion

This research demonstrates that for small medical datasets, a data-centric approach focusing on training dynamics is more effective than increasing architectural complexity. By flipping traditional training heuristics—extending burn-in and increasing regularization over time—it is possible to generate high-fidelity synthetic data that directly improves clinical classification tasks. Future work will explore multi-stage generative frameworks and broader applications in other medical domains.

Summary updated

BibTeX
@misc{pan2025r,
  author = {Tsung-Wei Pan and Chang-Hong Wu and Jung-Hua Wang and Ming-Jer Chen and Yu-Chiao Yi and Tsung-Hsien Lee},
  title = {{R3GAN-based Optimal Strategy for Augmenting Small Medical Dataset}},
  year = {2025},
  eprint = {2510.26828},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.26828},
}
Figure