Irregular Feature Enhancer for Low-dose CT Denoising

Abstract So far, deep learning based networks have been wildly applied in Low-Dose Computed Tomography (LDCT) image denoising. However, they usually adopt symmetric convolution to achieve regular feature extraction, but cannot effectively extract irregular features. Therefore, in this paper, an Irregular Feature Enhancer (IFE) focusing on effectively extracting irregular features is proposed by combining Symmetric-Asymmetric-Synergy Convolution Module (SASCM) with a hybrid loss module. Rather than simply stacking symmetric convolution layers used in traditional deep learning based networks, SASCM jointly utilizes symmetric and asymmetric convolution layers so as to effectively extract irregular tissue information of the image. In addition, the hybrid loss module is proposed to guide IFE to further mine the intrinsic feature information of the image from three perspectives: pixel point, high-level feature space, and gradient. The ablation experiments demonstrate the effectiveness and feasibility of SASCM and the hybrid loss. The quantitative experimental results also show that compared with several related LDCT denoising methods, the proposed IFE performs the best in terms of PSNR and SSIM. Furthermore, it can be observed from the qualitative visualization that the proposed IFE can recover the best image detail structure information among the compared methods.