ILKA-Net: CT Metal Artifact Reduction Network Based on Large Kernel Attention

Metal artifacts caused by metal objects in computed tomography (CT) greatly reduce the quality of CT images and affect the diagnosis of diseases for doctors. With the application of deep learning in medical imaging, various networks have been proposed to reduce CT metal artifacts. Because of the professionalism and privacy of medical imaging data, it is difficult to obtain large-scale data. Therefore, limited samples are commonly available, and lightweight networks are very suitable for processing small sample data. This paper proposes a U-Net based artifact reduction network named ILKA-Net, in which an Inception Large Kernel Attention (ILKA) module is proposed to reduce the number of parameters and improve the effect of reducing artifacts. In addition, large convolutional kernels are used instead of small convolutional kernels in the U-Net to achieve better effect of reducing artifacts.