AIACT-GAN: CT reconstruction based on dynamic attention and generative adversarial networks

Abstract

X-ray imaging is already a very mature technology. It is cheap and the radiation dose to the patient is very low. However, x-ray imaging can only provide two-dimensional information, not three-dimensional information of the patient's body. Computed Tomography (CT) can provide spatial information about the interior of the human body, giving the doctor more useful information, and the radiation dose to the patient is significantly higher. This is because conventional CT imaging techniques require a lot of X-rays for whole-body scanning. We introduce an end-to-end Generative Adversarial Network (GAN) network approach, AIACT-GAN, for the reconstruction of lung CT volumes directly from biplane x-ray images. In this work we reconstructed the CT in the presence of low radiation. We extracted features using a dynamic attention module and a dense connectivity module. In addition, in the fusion part we incorporated a contextual fusion module. The experimental results show that high quality CT can be reconstructed from x-ray images using AIACT-GAN.