Exploring the Robustness of Stable Diffusion in Chest X-ray Generation from Text Prompts

Medical imaging is essential for effective diagnosis, patient monitoring, and treatment planning. However, data scarcity, class imbalances, and concerns related to ethical and privacy limit the availability, accessibility and usage of image datasets in clinical settings, medical research, and training. Generative models, particularly Stable Diffusion, have emerged as promising options for generating high-quality medical images. The study investigates the stability of text prompting based Stable Diffusion which was trained on the NIH Dataset to generate clinically relevant chest X-ray images, using ClinicalBERT for domain-specific text encoding. To assess both clinical relevance and fault tolerance, a metric-driven approach is used: FID, GLCM, and KL divergence to evaluate clinical accuracy, while LPIPS, SSIM, and PSNR measure the reliability against variations in text prompts, including anatomical specificity, spelling and terminology, modality accuracy, and logical consistency. Our findings provide insights on text-conditioned Stable Diffusion model’s ability to produce clinically meaningful images while identifying significant possibilities for strengthening robustness, thereby contributing to advances in text-guided medical image synthesis.