Structural semantic enhancement network for low-dose CT denoising

Abstract

Most research on low-dose computed tomography primarily focuses on maximizing noise reduction, often at the expense of image structure and texture details. In this paper, we propose a Structural Semantic Enhancement Network (SSEN) that emphasizes the extraction and preservation of structural semantic features at different stages of the denoising process to enhance image sharpness. Specifically, unlike conventional methods that utilize a 3 × 3 Sobel operator for edge feature extraction, our approach employs a 5 × 5 Sobel operator with dense connections, preserving richer low-level semantics. Unlike conventional coordinate attention, which relies on 1 × 1 convolutional layers for feature activation, our approach employs 1 × 5 (or 5 × 1) asymmetric convolutional layers to expand the receptive field and capture richer global attention and contextual information. Furthermore, rather than commonly employed mean squared error loss functions, we propose a compound loss function that combines L₁ loss, multi-scale structural similarity index measure loss, and multi-scale perceptual loss, effectively recovering structural and perceptual features. This study indicates that the proposed method can effectively extract and utilize the structural semantic features to retain more image structure and texture details. In the experiments on the AAPM-Mayo Clinic LDCT Grand Challenge dataset, SSEN achieved a SSIM of 0.9193 and a PSNR of 33.6191, outperforming the comparison methods in terms of image quality restoration and structural information recovery.