Hybrid Perceptual Image Enhancement Framework using Traditional Processing and Deep Learning Refinement

Abstract

Medical image enhancement plays a crucial role in improving the visibility of anatomical structures and facilitating reliable visual interpretation of medical images. However, X-ray images may contain noise, inadequate contrast, and poorly distinguishable structural information, which can limit their interpretation accuracy. To address these limitations, we propose a framework that combines conventional preprocessing and contrast enhancement with DnCNN-based denoising, CNN feature extraction, and U-Net-based image reconstruction in a sequential refinement process. This integration aims to suppress unwanted artifacts while preserving clinically relevant structures and improving the overall perceptual quality of medical images. The original images are first resized and scaled to have a uniform representation along the entire processing chain. Median filtering is used to remove impulse noise while maintaining image boundary information. Contrast-limited adaptive histogram equalization (CLAHE) is then used to enhance local contrast, and unsharp masking is used to amplify high-frequency structural details. The pre-processed images are further provided to a Denoising Convolutional Neural Network (DnCNN) to remove residual noise using residual learning. The resulting denoised images are processed using a CNN feature extraction module to learn hierarchical representations associated with edges, textures, intensity variations, and anatomical structures. These learned representations were subsequently provided to a U-Net-based reconstruction and refinement network incorporating encoder-decoder processing and skip connections to preserve spatial details while performing multiscale enhancement. A composite objective incorporating reconstruction, perceptual, and edge preservation losses guided the refinement process to improve the image fidelity and structural consistency. The effectiveness of the proposed framework is evaluated using Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), Natural Image Quality Evaluator (NIQE), Normalized Absolute Error (NAE), Contrast-to-Noise Ratio (CNR), Mean Squared Error (MSE), Total Training Loss (Total Loss), and Lightness Order Error (LOE). The experimental results demonstrate that, compared with intermediate enhancement stages and original input images, the proposed sequential framework provides superior contrast enhancement, noise suppression, structural preservation, and perceptual quality while maintaining diagnostically relevant anatomical information.