Low-Dose CT Image Denoising with a Residual Multi-scale Feature Fusion Convolutional Neural Network and Enhanced Perceptual Loss

Computed tomography (CT) stands as a pivotal medical imaging technique, delivering timely and reliable clinical evaluations. Yet, its dependence on ionizing radiation raises health concerns. One mitigation strategy involves using reduced radiation for low-dose CT (LDCT) imaging; however, this often results in noise artifacts that undermine diagnostic precision. To address this issue, a distinctive CT image denoising technique has been introduced that utilizes deep neural networks to suppress image noise. This advanced CT image denoising network employs an attention mechanism for the feature extraction stage, facilitating the adaptive fusion of multi-scale local characteristics and channel-wide dependencies. Furthermore, a novel residual block has been incorporated, crafted to generate features with superior representational abilities, factoring in diverse spatial scales and eliminating redundant features. A unique loss function is also developed to optimize network parameters, focusing on preserving structural information by capturing high-frequency components and perceptually important details. Experimental results demonstrate the effectiveness of the proposed network in enhancing the quality of LDCT images.