Data-Driven Contrast-Enhanced Dual-Energy CT Imaging via Physically Constrained Attention

Computed tomography (CT) is widely used to generate cross-sectional views of the internal anatomy of a subject. Conventional CT imaging with single energy is, however, incapable of providing material composition information for various clinical applications because different materials may lead to the same CT numbers. Dual-energy CT (DECT) with physical means of simultaneously generating and measuring photon signals of two different spectra is designed to break this degeneracy. While valuable, this approach adds an extra layer of complexity on top of the widely used single-energy CT (SECT) and increases system costs, hindering the use of DECT scanners in less developed regions. Leveraging the ability of deep learning in nonlinear mapping and prior knowledge extraction from routine clinical data, here we develop a data-driven, lightweight strategy of obtaining DECT images from SECT images using a physically constrained attention mechanism. The proposed strategy is evaluated comprehensively by using high-fidelity simulation datasets and clinical contrast-enhanced DECT datasets. In terms of both prediction accuracy and inference speed, our method exhibits notable advantages over a variety of existing approaches. This technique holds the potential to provide a fast and cost-effective solution for contrast-enhanced spectral CT, catering to a broad range of CT applications.