Publication

Abstract : Noise-free X-ray imaging is crucial for accurate medical diagnostics. However, noise remains a persistent challenge, hindering diagnostic accuracy. This work introduces X-GAN, a physics-informed generative adversarial network (GAN) designed to denoise chest X-ray images while preserving critical anatomical structures. X-GAN integrates a U-Net generator with an edge-preserving attention mechanism and a PatchGAN discriminator to enhance image quality and retain fine anatomical details. The generator incorporates gradient information at its bottleneck layer to emphasize edges, thereby better preserving anatomical features. A hybrid loss function combines adversarial, reconstruction, and edge-aware losses to improve performance. The discriminator evaluates local regions of the generated images, providing feedback to refine denoising iteratively. The X-GAN framework is rigorously evaluated on two publicly available datasets for quantitative benchmarking, downstream diagnostic classification, and blinded mean opinion score (MOS) by two expert radiologists. The proposed X-GAN model with 2.17M parameters outperformed state-of-the-art (SOTA) deep learning (DL) models, including RIDNet, CBDNet, CGAN, and X-ReCNN. X-GAN achieved a peak signal-to-noise ratio (PSNR) of 36.48 dB, a structural similarity index (SSIM) of 0.95, and a superior edge preservation index (EPI) of 0.721. In downstream diagnostic classification tasks, denoising with X-GAN improved ResNet-50 accuracy from a baseline of 54.55% to 79.52%. Furthermore, blind clinical assessments by radiologists confirmed high diagnostic confidence. By integrating physics-informed edge constraints into a GAN framework, X-GAN achieved a better trade-off between suppressing quantum noise and retaining the diagnostic, important anatomical feature. The X-GAN significantly improved the reliability of downstream automated classification tasks and, due to its computational efficiency, X-GAN is suitable for integration into real-time automated diagnostic tools.