A Novel Attention-Guided Enhanced U-Net With Hybrid Edge-Preserving Structural Loss for Low-Dose CT Image Denoising

Abstract

Computed Tomography (CT) scan, pivotal for medical diagnostics, involves exposure to electromagnetic radiation, potentially elevating the risk of leukemia and cancer. Low-dose CT (LDCT) imaging has emerged to mitigate these risks, extensively reducing radiation exposure by up to 86%. However, it significantly reduces the quality of LDCT images and introduces noise and artifacts, degrading the diagnostic accuracy of the Computer Aided Diagnostic (CAD) system. This study presents a novel U-Net architecture, featuring several key enhancements. The model integrates residual blocks to improve feature representation and employs a custom hybrid loss function that combines structural loss with gradient regularization using the Euclidean norm, promoting superior CT image quality retention. Additionally, incorporating Attention Gates in the up-sampling layers of a proposed model optimizes the extraction of critical features, ensuring more precise denoising of CT images. The proposed model undergoes iterative training, using a custom loss function to refine its parameters and improve CT image denoising progressively. Its performance is rigorously evaluated both qualitatively and quantitatively on the ‘2016 Low-dose CT AAPM Grand Challenge dataset’. The results, assessed through the metrics Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Root Mean Square Error (RMSE), demonstrated promising improvements compared to state-of-the-art techniques. The model effectively reduces noise while preserving critical fine details, establishing itself as a highly efficient solution for LDCT image denoising.