Deep-learning Based Prediction of Virtual Non-contrast CT Images

Abstract

In this paper, we present a method, based on deep learning, for prediction of non-contrast CT image from a single contrast image. For training of this image-to-image translation task, virtual contrast and virtual non-contrast (VNC) images were created from spectral CT data by Philips IntelliSpace Portal (ISP) software. Virtual version of conventional CT (cCT) images and VNC images allows to train paired supervised image-to-image translation models. Two different schemes were tested to train the Convolutional Neural Network (CNN) with U-Net architecture, using standard training with L1/L2 loss as well as training via conditional Generative Adversarial Network (cGAN) with an additional Wasserstein modification (WcGAN). Qualitatively, the proposed method achieves similar results to the original VNC images. However, quantitatively, the trained CNN provides a slightly smaller density reduction in some tissues. Non-contrast image can be predicted from a single conventional CT image, without the need for pre- and post-contrast scan or without a spectral CT scan.