Hybrid Image Denoising via Sliced Ridgelet Transform and Attention-Guided Deep Learning for Medical and Industrial Applications

Image denoising is a critical task in medical diagnostics and industrial inspection, where noise can significantly degrade structural clarity and compromise analytical accuracy. Unlike conventional approaches that operate directly on noisy inputs, the proposed framework introduces transform-domain preconditioning to simplify the learning space and enhance structural representation prior to learning. The transform-domain stage employs the Sliced Ridgelet Transform to generate sparse, directionally structured representations using Radon projections and adaptive thresholding, preserving edges and structural information. The enhanced representations are refined using an attention-guided residual U-Net with multi-scale feature fusion and perceptual learning. An optional adversarial learning module with a PatchGAN discriminator is included to further improve visual realism. By conditioning the network on transform-domain outputs, the learning process focuses on residual noise suppression and fine texture reconstruction rather than coarse denoising. The proposed framework is evaluated on benchmark datasets, including BraTS 2021, NIH Chest X-ray, and industrial radiographic images, under multiple noise conditions such as Gaussian, speckle, and mixed noise. Experimental results demonstrate consistent and significant improvements over state-of-the-art methods, including DnCNN and SwinIR, achieving higher peak signal-to-noise ratio and structural similarity values along with reduced perceptual error metrics such as LPIPS and FID. The results indicate that the integration of transform-domain sparsity with deep learning-based refinement provides a robust and generalizable solution for complex noise removal, structural preservation, and high-fidelity texture restoration across diverse imaging domains. The proposed framework provides an effective balance between noise suppression, structural preservation, and perceptual quality, making it suitable for real-world medical and industrial imaging applications.