Chest X-Ray Image Denoising Using Edge-Enhanced U-Net and Generative Adversarial Networks

Abstract

Chest X-ray image quality is crucial for accurate medical diagnosis, yet acquired images are often degraded by noise, which can obscure important anatomical details. This work presents a deep learning–based denoising framework for chest X-ray images using a customized U-Net-like architecture combined with adversarial learning. Normal chest X-ray images from the Kaggle Chest X-ray Pneumonia dataset are preprocessed, resized, and synthetically corrupted with Gaussian noise to generate paired noisy and clean training samples. The proposed generator network, termed SharpXRLikeUNet, integrates Laplacian filtering to enhance edge preservation during denoising. Initially, the model is trained using a composite loss function consisting of L1 loss, structural similarity index measure (SSIM) loss, and perceptual loss derived from a pretrained VGG16 network. To further improve visual realism, a generative adversarial network (GAN) framework is employed by introducing a patch-based discriminator, and an adversarial loss component is incorporated into the generator objective. Experimental results demonstrate effective noise reduction while preserving fine structural details in chest X-ray images. The trained models are saved, and the generator is exported to ONNX format to support efficient deployment and cross-platform inference.

Introduction

The text presents a deep learning–based framework for denoising chest X-ray images, addressing challenges posed by noise during acquisition, transmission, or low-dose imaging. Traditional denoising methods often over-smooth images, losing critical anatomical details. To overcome this, the proposed approach uses a customized U-Net–like generator, SharpXRLikeUNet, enhanced with Laplacian filtering for edge preservation, and trained with a composite loss function combining L1, SSIM, and perceptual losses.

To further improve realism, a patch-based GAN discriminator is employed, encouraging the generator to produce visually accurate and sharp outputs. The framework is trained on the Kaggle Chest X-ray Pneumonia dataset with simulated Gaussian noise and evaluated qualitatively for noise suppression and structural preservation.

The system includes an ONNX-exported generator for cross-platform deployment and a React-based frontend for interactive demonstration, enabling real-time denoising of uploaded X-ray images. Expected results indicate significant improvements in visual clarity, edge sharpness, and preservation of fine anatomical structures, making the framework suitable for clinical and research applications.

Conclusion

This work presented a deep learning–based framework for denoising chest X-ray images using a customized U-Net-like architecture combined with adversarial learning. The proposed SharpXRLikeUNet integrates Laplacian-based edge enhancement to preserve fine anatomical details while effectively suppressing noise. The model was trained in two stages, initially using a composite reconstruction loss consisting of L1, SSIM, and perceptual loss, followed by adversarial refinement using a patch-based discriminator. Experimental results demonstrate stable training behavior and effective convergence, with low training and validation losses achieved during the reconstruction phase and balanced generator–discriminator losses during adversarial training. Qualitative evaluation shows that the GAN-enhanced model produces denoised outputs with improved sharpness and structural fidelity compared to non-adversarial reconstruction alone. These findings indicate that the proposed approach successfully balances noise reduction and detail preservation, making it suitable for medical image enhancement applications.

References

[1] K. Dabov, A. Foi, V. Katkovnik, and K. Egiazarian, “Image denoising by sparse 3-D transform-domain collaborative filtering,” IEEE Transactions on Image Processing, vol. 16, no. 8, pp. 2080–2095, Aug. 2007. [2] A. Buades, B. Coll, and J.-M. Morel, “A non-local algorithm for image denoising,” in Proc. IEEE Computer Society Conf. Computer Vision and Pattern Recognition (CVPR), 2005, pp. 60–65. [3] O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional networks for biomedical image segmentation,” in Proc. Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2015, pp. 234–241. [4] C. Ledig et al., “Photo-realistic single image super-resolution using a generative adversarial network,” in Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2017, pp. 4681–4690. [5] J. Johnson, A. Alahi, and L. Fei-Fei, “Perceptual losses for real-time style transfer and super-resolution,” in Proc. European Conf. Computer Vision (ECCV), 2016, pp. 694–711. [6] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: From error visibility to structural similarity,” IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600–612, Apr. 2004. [7] I. Goodfellow et al., “Generative adversarial nets,” in Proc. Advances in Neural Information Processing Systems (NeurIPS), 2014, pp. 2672–2680. [8] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros, “Image-to-image translation with conditional adversarial networks,” in Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2017, pp. 1125–1134. [9] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778. [10] G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Densely connected convolutional networks,” in Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2017, pp. 4700–4708.

Copyright

Copyright © 2026 Danish Wilson, Alsa Elizabeth Sali , Anna SP, Athulya M, Aneesa HA. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.