CASPER: Conventional CT database Augmentation using deep learning based SPEctral CT images geneRation

Spectral CT scanner is an emerging technology in the clinical field, which al-lows to generate from a single scan conventional HU images, virtual non-contrast images (VNC) corresponding to unenhanced CT images on conventional CT scanners, and virtual mono-energetic (monoE) images at different keV that mimics low (at high keV) to high (at low keV) iodine-based contrast-enhanced studies. It has been demonstrated that these spectral images could be used as data augmentation for cardiovascular structures segmentation, improving dice scores on contrasted images, and enabling to segment true non-contrast (TNC) conventional scans. However, as of now, spectral CT are not widely distributed, making this data augmentation process not widely available.To overcome this limitation, we propose in this study CASPER, a full data augmentation workflow to provide any conventional CT training dataset with spectral CT augmentation capabilities. We trained on an unannotated data-base of 2500 spectral scans an image translation network (HUSpecNet) to create monoEs and VNC from conventional HU images. Any conventional CT scan exam can then be translated to mimic different contrast agent injection protocols.We evaluated our CASPER methodology by training a 3D U-net to segment the aorta, both with and without data augmentation on 143 annotated conventional scans. The network performances were compared on a testing dataset of 70 patients, including pulmonary embolism (PE) early enhancement scans, full chest-abdominal-pelvic delayed enhancement scans and TNC scans. The network trained using CASPER outperformed significantly the network trained without it, on every imaging protocols, and especially on TNC.