DyeFreeNet: Deep Virtual Contrast CT Synthesis

Abstract

To highlight structures such as blood vessels and tissues for clinical diagnosis, veins are often infused with contrast agents to obtain contrast-enhanced CT scans. In this paper, the use of a deep learning-based framework, DyeFreeNet, to generate virtual contrast abdominal and pelvic CT images based on the original non-contrast CT images is presented. First, to solve the overfitting issue for a deep learning-based method on small datasets, a pretrained model is obtained through a novel self-supervised feature learning network, whereby the network extracted intensity features from a large-scale, publicly available dataset without the use of annotations and classified four transformed intensity levels. Second, an enhanced high-resolution “primary learning generative adversarial network (GAN)” is then used to learn intensity variations between contrast and non-contrast CT images as well as retain high-resolution representations to yield virtual contrast CT images. Then, to reduce GAN training instability, an “intensity refinement GAN” using a novel cascade intensity refinement strategy is applied to obtain more detailed and accurate intensity variations to yield the final predicted virtual contrast CT images. The generated virtual contrast CTs by the proposed framework directly from non-contrast CTs are quite realistic with the virtual enhancement of the major arterial structures. To the best of our knowledge, this is the first work to synthesize virtual contrast-enhanced abdominal and pelvic CT images from non-contrast CT scans.