Dynamic Perception-oriented Low-dose CT Image Denoising Network using Structure-aware Self-similarity

Denoising low-dose computed tomography (CT) images is a challenging task in medical data processing. In this context, generative adversarial networks (GANs) have been employed to solve LDCT denoising problems. However, recent studies have revealed that basic GAN-based training introduces structural distortion. Generally, human perception of CT images is restored by a pre-trained network, including VGG-based perceptual loss. The extracted perceptual features lose CT-specific knowledge because the pre-trained networks are trained on natural images. In this context, we develop a Wasserstein GAN (WGAN) based denoising technique with three novel technical contributions to address the aforementioned issues. First, we introduce a novel Dynamic Convolution (DyConv) to build a generator network. Multiple parallel convolution kernels are aggregated dynamically using DyConv, based on input-dependent kernel attentions. Hence, it enhances model representation power without demanding the stacking of several 2D convolution layers. We propose a novel structure-aware network (SANet) trained on neighborhood structural input vectors via self-supervised contrastive learning. We integrate SANet with a neural network-based computational efficient non-local module (CE-NLM) to increase structural similarity. Furthermore, we propose a CT-specific perceptual loss computed on the perceptual feature space after the pre-trained network is fine-tuned to the CT data via transfer learning. The proposed approach demonstrates a superior performance as compared to state-of-the-art techniques. The new method is validated qualitatively and quantitatively on the two publicly accessible "2016 NIH-AAPM-Mayo Clinic Low Dose CT Grand Challenge" and the "Low-dose CT image and projection" dataset.