Synthetic Medical Imaging with Pathology-Aware Variational Autoencoders

Keywords: medical image synthesis, pathology-aware generation, variational autoencoder, synthetic data augmentation, chest X-ray, low-label learning, calibration (ECE), robustness to distribution shift, conditional prior, feature preservation

TL;DR: A pathology-aware VAE synthesizes class-consistent chest X-rays that raise detector AUC from 0.715→0.822 with only 10% labels, improving sensitivity and calibration.

Abstract:

Medical image analysis often faces severe label scarcity and privacy constraints.
Wepresent a Pathology-Aware Variational Autoencoder (PA-VAE) that prioritizes
preservation of clinically salient features during synthesis via a feature-matching
loss and a class-conditional latent prior. Using public chest radiograph settings
with low-label regimes (10clinical utility. On a simulated but reproducible bench
mark, PA-VAE improves downstream classification AUC from 0.715 (real-only)
to 0.822 with higher sensitivity at 95(0.091→0.295) and reduced calibration error
(0.017→0.026). The generator achieves competitive fidelity (lower FID-like) and
reconstruction quality (SSIM), and ablations indicate the feature- preservation loss
and class-conditional prior as principal contributors. Robustness analyses show
moderate degradation under adversarial-like and temporal drift perturbations. We
release a dependency-light, fully reproducible pipeline that procedurally synthe
sizes data, regenerates all figures, and exports JSON metrics to facilitate transpar
ent evaluation and future extensions.

Submission Number: 249