UNeXt: a Low-Dose CT denoising UNet model with the modified ConvNeXt block
Farzan Niknejad Mazandarani, Paul Babyn, Javad Alirezaie
Abstract
In recent decades, clinicians have widely utilized computed tomography (CT) for medical diagnosis. Medical radiation is potentially hazardous and therefore reducing x-ray radiation in CT scanning is desired. However, decreasing radiation dose leads to increased noise and artifacts. In this paper, low-dose CT images (LDCT) have been denoised in the UNet-based novel architecture of convolutional neural network (CNN) and compared with normal-dose images (NDCT). A multi-feature extraction block (MFEB) is placed to get extra features in the different receptive fields. The modified ConvNeXt block for CT images (CTNeXt) is developed to extract diverse feature data at various scales. Furthermore, we introduced the image reconstruction block to gradually merge the group convolutions’ feature information and eliminate the gap between the features to ease the transmission of multi-scale information from subsequent stages. The network is optimized using the integration of mean-squared error (MSE), mean-absolute error (MAE), and contrastive loss via vgg16-net. These functions show that they could effectively prevent edge over-smoothing, improve image texture, and preserve structural details. A comparative analysis of the proposed network demonstrates that our method outperforms state-of-the-art denoising models, such as Wasserstein Generative Adversarial Network (WGAN-vgg) and Residual Convolutional Encoder-Decoder (RED-CNN).