Conference Agenda

Medical image generation can support the development of robust deep learning models by providing controlled synthetic data when real datasets are limited, imbalanced, or difficult to share. This study presents a preliminary pipeline for conditional chest X-ray generation using a class-conditioned Latent Diffusion Model. The model was trained on chest radiographs labelled as No Finding, meaning that no pathological finding is reported in the dataset metadata, and conditioned on radiographic view position, namely PA and AP projection. Images were encoded into a latent space using a frozen Stable Diffusion VAE, while a MONAI Diffusion Model UNet was trained to denoise latent representations conditioned on the target view class. Several configurations were evaluated by varying training epochs, conditioning tokens, guidance scale, inference steps, and AP oversampling. Generated images were assessed using FID for realism, MS-SSIM for diversity, and auxiliary ResNet18 and DenseNet121 classifiers for conditioning fidelity. The best results showed a substantial improvement in FID, especially for AP images, while classifier evaluation confirmed that generated samples were consistently recognized as intended PA or AP class. The results suggest that class-conditioned latent diffusion is a promising approach for controlled chest radiograph synthesis, although future work should include chest-specific autoencoders and anatomy-aware evaluation.