Edge-guided ControlNet synthetic chest x-ray augmentation and ensemble classification

Chest X-ray (CXR) analysis is essential for the early detection of thoracic diseases, but challenges like limited data and class imbalance can undermine model performance. This study introduces a two-phase framework that integrates synthetic image generation with classification. The first phase utilizes a ControlNet-based Stable Diffusion pipeline, utilizing Canny–Sobel edge information to produce realistic synthetic CXRs. Features are extracted using ViT, Swin, ResNet, and handcrafted Gray-Level Cooccurrences Matrix (GLCM) descriptors. In the second phase, the enhanced dataset is classified using ensemble methods, where Extra Tree demonstrates superior effectiveness. Tests on the IU X-Ray and COVID-19 datasets show the framework’s success, achieving SSIM values of 89.53 for COVID and 86.21 for IU X-Ray, along with classification accuracies of 95.01% for COVID and 78.76% for IU X-Ray. These findings underscore the framework’s ability to mitigate data limitations and improve prediction accuracy, suggesting its viability for clinical applications.