Spectral CT Based Training Dataset Generation and Augmentation for Conventional CT Vascular Segmentation

Abstract

Deep learning has proved to be a very efficient tool for organs automated segmentation in CT scans. However, variation of iodine contrast agent concentration within the vascular system or organs is a major source of variation in image contrast. This requires building large databases representative of the important differences in contrast enhancement across CT studies. Furthermore, creating a low- or non-enhanced annotated database is still a very laborious task as semi-automatic segmentation software and even expert eyes often fail to find structures’ edges on low contrast images.

In this study, we aim to develop a new deep-learning network training approach based on spectral data augmentation using dual energy spectral CT (Philips iQon) images as training dataset. Indeed this new generation of CT scanners allows generating, from a single scan, virtual non-contrast images (VNC), corresponding to unenhanced CT images on conventional CT scanners, and virtual mono-energetic (monoE) images at different kV, that mimics low (at high kV) to high (at low kV) iodine-based contrast-enhanced studies. An experienced radiologist can then segment the target structures on highly contrasted low kV monoE with a semi-automatic tool (ISP, Philips) yielding ground truth for both monoE and VNC images.

As an illustration, we trained a 3D U-net convolutional neural network for aorta segmentation on conventional CT images. In addition to greatly facilitate the creation of an annotated non-contrast CT database, we demonstrate through multiple training experiments that using a variable proportion of VNC and randomly chosen monoE kV levels during the data augmentation process allows training networks that are able to segment target structures regardless of their level of contrast enhancement.