Development of Conditional Latent Diffusion Models for Pediatric Chest X-Ray Generation and Counterfactual Image Editing

Medical imaging plays a central role in modern healthcare, enabling non-invasive visualization for diagnosis and treatment planning. However, the development of robust diagnostic AI is hampered by limited, heterogeneous, and privacy-sensitive datasets. This challenge is acute in pediatric chest radiography, a critical sector where data curation is complicated by age-dependent anatomy, motion artifacts, and strict dose-reduction protocols. To address these limitations, Generative Artificial Intelligence (GenAI) has emerged as a transformative paradigm, offering the potential to synthesize high-fidelity, privacy-compliant medical data that can augment limited cohorts and correct distributional imbalances. These models also enable the visualization of how specific pathologies would manifest or evolve on patient-specific anatomy, providing a powerful tool for diagnostic explainability and comparative analysis. Building on this premise, the objective of this work is to develop and evaluate a conditional Latent Diffusion Model (LDM) for pediatric chest X-ray synthesis, enabling both pathology conditional generation and counterfactual editing to support diagnostic AI and clinical education. Methodologically, the study leverages the Medical Open Network for AI (MONAI) framework to implement autoencoder-based latent compression and a diffusion U-Net trained on PneumoniaMNIST dataset, with conditioning on clinical status, Bayesian hyperparameter optimization using Optuna, and advanced training strategies such as Exponential Moving Average and Classifier-Free Guidance for stability and semantic control. Quantitatively, the proposed pediatric LDM achieves a Fréchet Inception Distance (FID) of 66.3. This performance is competitive with state-of-the-art systems trained on significantly larger adult cohorts and markedly outperforms non-domain-adapted baselines. The model's generalization capabilities were further corroborated through extensive evaluation on the external VinDr-PCXR dataset, confirming its robustness in generating plausible pediatric anatomy beyond the training distribution. Qualitatively, the model produces anatomically consistent radiographs with high label fidelity. The diagnostic validity of both the synthetic cohorts and the counterfactual edits was confirmed through a fine-tuned classifier, and expert radiologist assessment via structured questionnaires, demonstrating the model's capacity for precise semantic synthesis and identity-preserving modifications. These results demonstrate the feasibility of controlled LDMs for pediatric chest imaging under limited data regimes and highlight their potential for privacy-preserving data augmentation, diagnostic model stress testing, and scalable, reusable pipelines for future multi-pathology and multi-modality clinical applications.