Controllable synthesis of abnormal chest X-ray images based on diffusion model

• 1、School of Intelligent Engineering and Automation, Beijing University of Posts and Telecommunications, Beijing 100876

Abstract：To address the scarcity of abnormal chest X-ray samples and the high cost of pixel-level annotations, we propose a controllable anomaly synthesis and data augmentation method under the constraint of anatomical consistency. Leveraging a text-conditioned diffusion model fine-tuned on medical image-text pairs, we use lung and cardiac masks as structural conditions to guide the generation process, enhancing the anatomical stability and usability of the synthesized images. Furthermore, pixel-level differences between normal and generated abnormal images are automatically used to form anomaly masks, enabling data augmentation for improved anomaly detection model training. Experiments show that anatomical structure constraints significantly improve anatomical consistency and the reliability of pseudo-labels. Evaluated on the ChestX-Det dataset for anomaly detection, incorporating synthetic abnormal data augmentation leads to a Pixel\_AUC improvement of 1.68\% and an Image\_AUC improvement of 0.2\%, demonstrating that the proposed method can effectively enhance anomaly detection performance without requiring additional manual annotations.

Keywords： Abnormal Chest X-ray Detection Controllable Abnormality Synthesis Diffusion Models Structure-Conditioned Guidance