Voxel-Sensitive Attention-Based Lattice Structure Residual Network for Low-Dose CT Image Denoising Using Wavelet-Based Structure Loss

Low-dose computed tomography (LDCT) imaging safeguards patients from potentially harmful X-rays. However, the signal-to-noise ratio (SNR) and computed tomography (CT) image quality deteriorate due to lower X-ray flux in the LDCT imaging modality. The custom design of deep networks with vanilla convolution layers ignores sensitive voxel information. Hence, voxel details for reconstructing low-contrast regions and high-frequency lesion edges may be lost during noise suppression. In addition, existing techniques do not satisfactorily address the problem of adversarial learning-induced structural deformation. This article proposes a LDCT denoising method based on Wasserstein GAN (WGAN) with three novel contributions. First, we introduce a novel build block called a voxel-sensitive attention-based lattice residual block (VSA-LRB). The generator network with stacked VSA-LRB can preserve the long-range pixel dependency while aggregating the voxel-sensitive features extracted along the spatial ( $H \times W$ ) and depthwise (D) directions. Furthermore, we train the proposed denoising network to learn efficient structural features via our proposed wavelet-based structural loss. The error is estimated on the wavelet-domain structural feature space, extracted by a dedicated pretrained CT-specific structure-aware network (CT-SANet). Hence, the wavelet-based loss function effectively mitigates structural distortions. We introduce a differential content loss to improve high-frequency lesion boundaries. The proposed method is validated qualitatively and quantitatively using two publicly available datasets: the “2016 NIH-AAPM-Mayo Clinic LDCT Grand Challenge dataset” and the “Low-dose-CT Image and Projection dataset.” The proposed method demonstrates a superior performance when compared with state-of-the-art techniques.