UNet with ResNextify and IB modules for low-dose CT image denoising

In recent years, low-dose computed tomography (LDCT) scans have surpassed traditional CT scans in popularity as people grow more health conscious. However, the resulting CT images are full of noise and artifacts, therefore a growing number of researchers are trying to figure out how to make better images. Recently, there has been a lot of research done on deep learning to eliminate artifacts in low-dose computed tomography (LDCT). But in comparison to traditional denoising techniques, it performs better thanks to data-driven execution and fast performance. The majority of the recently suggested UNet based approaches, however, have issues with residual noise, over-smoothed structures and leads to more and more complex networks. Thus, we have proposed a new approach that combines the per-pixel feedback capability of the U-Net architecture with the ResNextify and inverted bottleneck (IB) from ConvNeXt model to enhance the denoising network. One of the two sub-networks in this novel generator, processes the decomposed high-frequency components of an LDCT picture. Data from the entire LDCT image is processed using a different one. When tested on a publicly available dataset, experimental results clearly show that the proposed approach surpasses other approaches such as BM3D, K-SVD, CCADN, CycleGAN, and SKFCycleGAN in terms of protecting structure information with reducing noises at satisfactory level. This is evident from the model's ability to achieve the highest values for PSNR and SSIM. This research paper aims to elucidate the UNet embedded with Resnextify and inverted bottleneck modules for CT image denoising. © Bharati Vidyapeeth's Institute of Computer Applications and Management 2024.