Synthetic data augmentation fails to improve CT emphysema subtype classification in small-data study

No synthetic data augmentation method (cVAE, cSD-VAE, cWGAN-GP) significantly improved balanced accuracy for CT-based emphysema subtype classification over baseline (0.522 ± 0.067) in 168 patches. Best method (cWGAN-GP at 1.0x) achieved 0.548 ± 0.068 (p=0.084).

• No generative augmentation strategy (cVAE, cSD-VAE, cWGAN-GP) or classical geometric augmentation significantly improved balanced accuracy over the unaugmented baseline.
• The best-performing method, cWGAN-GP at 1.0x synthetic-to-real ratio, achieved a balanced accuracy of 0.548 ± 0.068 (Cohen's d=0.47, p=0.084).
• Pixel-space and feature-space distributional fidelity did not correlate with downstream classification performance.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.