Synthesizing and Evaluating Chest X-Ray Images Using LoRA-Adapted Stable Diffusion: A Study on Task-Dependent Utility

This study investigates the utility of synthetic pediatric chest X-ray images, generated by a Stable Diffusion model fine-tuned with Low-Rank Adaptation (LoRA), based on the complexity of the diagnostic task. We evaluated the generated data not only with standard metrics like FID and CLIP Score but also through its application in both a fine-grained 3-class (Normal, Bacterial, Viral Pneumonia) and a coarse-grained 2-class (Normal, Pneumonia) downstream classification task. Our results reveal a significant contradiction: the same synthetic data that was detrimental to the 3-class model’s performance (accuracy drop from 60% to 57%) acted as an effective data augmentation tool for the 2-class task, significantly improving its accuracy from 76% to 83%. These findings demonstrate that the utility of synthetic data is highly task-dependent and that downstream task-based evaluation is essential for validating the true clinical value of synthetic medical images beyond what standard fidelity metrics can provide.