Attenuation image referenced (AIR) effective atom number image calculation for MeV dual-energy container CT using image-domain deep learning framework

Abstract

Traditional inspection technology for cargo or container imaging in customs and harbours is MeV X-ray radiography. The biggest limitation for this technology is the structural overlapping problem, which is inherent to radiography technology. MeV dual energy CT has a major advantage over radiography in that it can provide cross-section image, which is free of the structural overlapping problem. Besides, the recorded dual-energy projection data provides the ability for material decomposition. Electron density image and effective atom number image can be further calculated from the material decomposition coefficient images. However, the quality of effective atom number image can be very poor. The behind reasons are multifaceted. In this paper we proposed an Attenuation Image Referenced (AIR) effective atom number image calculation method for MeV dual-energy container CT imaging by using an image-domain neural network. The network has three channels as input and outputs with the estimated effective atom number image. The input three channels include the low and high-energy attenuation images and the effective atom number image that was directly calculated by using the derived formula. The network utilizes the low and high-energy attenuation image as guidance or reference for the restoration of effective atom number image. The network was trained on synthetic data, which is based on the shape of XCAT model but filled with materials that often appear in security imaging. The trained network also performed well on experimental data, showing the robustness and good generalization ability of the network. The quantitative analysis on the simulation and experimental data that comes from actual MeV dual-energy CT system showed the effectiveness of the proposed Attenuation Image Referenced (AIR) deep learning method for effective atom number image calculation.