Exploiting Computational-Efficient Hybrid-Scale Self-Similarity for Low-Dose CT Image Denoising

Data-driven low-dose computed tomography (LDCT) denoising methods using generative adversarial networks (GANs) have recently gained popularity. However, the deep network (DN) design with a stack of vanilla convolutional layers ignores essential global voxel details and is inefficient in improving the signal-to-noise ratio. Recent studies reveal that GAN-based training may introduce structural deformation. Many state-of-the-art LDCT denoising techniques exploit highly computationally expensive self-similarity via neural network-based nonlocal modules (NLMs). In this article, we mitigate these issues using a rectified-Wasserstein GAN (ReWGAN)-based LDCT noise reduction technique. We develop a generator network which is an ensemble of DN and shallow networks (SNs). The DN architecture contains a stack of inventive multiscale position-sensitive attention-based residual blocks (MSPSA-RBs) and a computationally efficient hybrid-scale self-similarity exploitation network (HS-SENet). MSPSA-RB efficiently extracts spatially sensitive multiscale features to preserve the local and global pixel dependency of CT images. HS-SENet exploits single- and cross-scale similarity simultaneously improving the noise-suppression capability of the generator network. Moreover, we fuse the shallow features via SN, which can expedite the network’s convergence and help restoring the structural features of CT images. In addition, we propose a novel CT-specific structure-preserving loss (CT-SPL) that can overcome the limitations of conventional perceptual loss functions, including cross-hatch artifacts and structural distortion. Extensive numerical experiments are conducted to validate the proposed method qualitatively and quantitatively on the “2016 NIH AAPM-Mayo Clinic dataset” and the “LDCT projection dataset.” The new method demonstrates superior performance in terms of the peak signal-to-noise ratio (PSNR = 35.9901 dB), structural similarity (SSIM = 0.9627), and visual information fidelity (VIF = 0.2075) compared with state-of-the-art techniques.