An efficient approach to synthesize medical images for data augmentation with constrained training data

Abstract

Generative Adversarial Networks (GANs) have achieved notable success in data synthesis; however, their performance degrades substantially when training data are scarce. In this work, we propose a novel augmentation and regularization framework designed to improve GAN training stability and generation quality under limited-data conditions. The proposed approach consistently outperforms existing state-of-the-art methods across multiple GAN architectures, yielding improved Inception Score (IS) and reduced Fréchet Inception Distance (FID) on diverse benchmark datasets. We further demonstrate the effectiveness of the proposed framework in medical image synthesis, where synthetic data augmentation is used to enhance downstream segmentation performance. Experimental results on multiple medical imaging datasets show that models trained with the proposed augmentation strategy achieve up to a 3–5% improvement in Dice score compared to standard augmentation techniques. In data-constrained settings, the method also produces images with improved visual fidelity and lower FID relative to existing GAN-based baselines. These results highlight the potential of the proposed framework as a reliable data augmentation strategy for training deep models under constrained data availability. Future work will extend the approach to multi-category and instance-level synthesis with lesion-specific annotations.