Nicholas Dietrich, David McShannon
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).
Introduction
Chronic obstructive pulmonary disease (COPD) is a leading cause of morbidity and mortality worldwide, and emphysema, one of its principal components, is characterized by irreversible destruction of lung parenchyma [1]. Computed tomography (CT) allows for direct visualization of emphysematous changes, and texture-based analysis of CT image patches has become a widely used strategy for characterizing emphysema subtypes, including normal tissue (NT), centrilobular emphysema (CLE), and paraseptal emphysema (PSE) [2].
Materials and methods
Study design
This was a retrospective computational study evaluating synthetic data augmentation strategies for a three-class emphysema subtype classification task. This study was designed and reported in accordance with the Checklist for Artificial Intelligence in Medical Imaging (CLAIM) (S1 Checklist). The study used a publicly available benchmark dataset which did not involve new human data, or data that could identify individual patients during or after data collection [16]. All code was executed in a stable environment using a single NVIDIA A100 GPU.
Results
Overall classification performance
Table 1 presents the primary performance metrics across all 14 experimental conditions. The baseline classifier (pretrained ResNet18, no augmentation) achieved a mean balanced accuracy of 0.522 ± 0.067 across 10 seeds. No individual augmentation strategy produced a statistically significant difference in balanced accuracy relative to baseline (Wilcoxon signed-rank test p-values > 0.05). Detailed per-seed results for all methods are provided in S2 Table.
Discussion
This study evaluated whether synthetic data augmentation using three class-conditional generative approaches or classical geometric transformations improves emphysema subtype classification when training data is limited. Our principal finding was that no individual augmentation strategy produced a statistically significant improvement in balanced accuracy over the unaugmented baseline. These findings are relevant to clinical imaging research where labeled datasets are often limited and texture-based distinctions are critical, such as interstitial lung disease or emphysema phenotyping.
Conclusions
In a controlled comparison on a small emphysema CT patch dataset, generative and classical geometric augmentation did not produce statistically significant improvements in emphysema subtype classification when applied individually or with a pre-specified ensemble combining all four augmentation methods. These findings suggest there may be limited standalone utility of generative augmentation in small-data, texture-driven tasks. Future work should focus on improving generation under small-data conditions using task-aware and pathology-constrained strategies.
Citation: Dietrich N, McShannon D (2026) Synthetic data augmentation for CT-based emphysema subtype classification: A comparative evaluation of generative and classical approaches. PLoS One 21(8): e0355850. https://doi.org/10.1371/journal.pone.0355850
Editor: Zeheng Wang, Commonwealth Scientific and Industrial Research Organisation, AUSTRALIA
Received: March 23, 2026; Accepted: July 27, 2026; Published: August 20, 2026
Copyright: © 2026 Dietrich, McShannon. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The Computed Tomography Emphysema Database analyzed in this study is publicly available from the original authors’ database website at https://lauge-soerensen.github.io/emphysema-database/. The database is provided free of charge for research and educational purposes. Questions regarding access to the original CT emphysema database should be directed to Lauge Sørensen at lauges@di.ku.dk, as specified on the database website. All source code, the completed CLAIM checklist, and the per-seed result files required to reproduce the analyses, tables, and figures are publicly available on GitHub (https://github.com/radiology-generative/emphysema-synthetic-augmentation) and archived on Zenodo with a permanent identifier (https://doi.org/10.5281/zenodo.20754048).
Funding: The author(s) received no specific funding for this work.
Competing interests: The authors have declared that no competing interests exist.
