Synthetic data augmentation for CT-based emphysema subtype classification: A comparative evaluation of generative and classical approaches.

Abstract

Data scarcity is a persistent challenge in medical image analysis. Synthetic data generation using deep generative models has been proposed as a potential approach to address this limitation, yet its performance in small-data settings remains poorly characterized. This study compared three class-conditional generative approaches, a conditional variational autoencoder (cVAE), a conditional shallow-decoder VAE variant (cSD-VAE), and a conditional Wasserstein GAN with gradient penalty (cWGAN-GP), against classical geometric augmentation for emphysema subtype classification on 168 CT patches (three classes: normal tissue, centrilobular emphysema, and paraseptal emphysema). Each method was evaluated at three synthetic-to-real ratios (0.5x, 1.0x, 2.0x) using patient-level 70/30 splits across 10 random seeds, with an ImageNet-pretrained ResNet18 as the downstream classifier. No individual augmentation strategy produced a statistically significant improvement in balanced accuracy over the unaugmented baseline (0.522 ± 0.067). The conditional WGAN-GP at 1.0x achieved the highest individual balanced accuracy (0.548 ± 0.068, Cohen's d = 0.47 versus baseline), but did not reach statistical significance (p = 0.084). A pre-specified ensemble combining all four augmentation methods at the 1.0x multiplier did not significantly improve balanced accuracy over baseline (0.541 ± 0.088 versus 0.522 ± 0.067; Cohen's d = 0.32, Wilcoxon p = 0.275). Neither pixel-space nor feature-space distributional fidelity was associated with downstream classification performance. Overall, no benefit was detected from class-conditional generative augmentation in this small-data, texture-driven setting. Future work should focus on improving generative modeling under small-data conditions, including task-aware objectives and pathology-constrained synthesis.