A Hybrid U-Net Framework for Low-Dose CT Image Denoising: Leveraging EdgeNet+ for Structural Integrity and Image Quality

Abstract

Computed Tomography (CT) is a non-invasive imaging modality widely used for the precise detection of abnormalities within the human body. However, the electromagnetic radiation generated during CT scans poses health risks, including metabolic disturbances and genetic mutations, which can increase cancer susceptibility. To mitigate these hazards, Low-Dose CT (LDCT) techniques were introduced, significantly reducing radiation exposure. However, LDCT compromises image quality by introducing increased noise, artifacts, reduced contrast and structural distortions, which can impair the accuracy and reliability of Computer-Aided Diagnosis (CAD) systems. This study presents EdgeNet+, an advanced U-Net architecture with 21 convolutional layers and three skip connections, which facilitate the maintenance of high-quality structural details by allowing features from earlier layers to directly influence later stages. EdgeNet+ enhances performance by incorporating a multi-level edge detection block and a hybrid loss function. The edge detection block effectively captures edge features across various scales. The hybrid loss function combines Structural Similarity Index Measure (SSIM) and L1 losses, where SSIM promotes the preservation of image structures and perceptual quality, while L1 loss ensures accurate pixel-wise reconstruction. This combination enables EdgeNet+ to produce denoised images that are visually appealing and faithful to the original content, making it highly effective in enhancing detection accuracy. In the ablation study, the impact of each component of the EdgeNet+ model was systematically assessed, demonstrating a significant improvement in denoising performance. Comparative analysis reveals a noteworthy 45.12% increase in Peak Signal-to-Noise Ratio (PSNR), a 26.17% enhancement in SSIM and a remarkable 90.71% reduction in Root Mean Square Error (RMSE) when juxtaposed with the LDCT image. Compared to benchmark algorithms, the proposed approach demonstrates a marked improvement in noise reduction and artifact removal. Qualitative comparisons reveal a high similarity between the denoised CT images produced by EdgeNet+ and normal-dose CT images