X-ray Monochromatic Imaging from Single-spectrum CT via Machine Learning

TL;DR

This paper introduces a machine learning approach using ResNet to generate monochromatic images from single-spectrum CT scans, reducing the need for dual-energy scans and associated costs.

Contribution

A novel deep learning method that reconstructs monochromatic images from single-spectrum CT data, bypassing the need for dual-energy scans.

Findings

Achieved less than 0.2% relative error in monochromatic image reconstruction.

Demonstrated high-quality images on testing data.

Potential to improve clinical DECT applications.

Abstract

In clinical CT system, the x-ray tube emits polychromatic x-rays, and the x-ray detectors operate in the current-integrating mode. This physical process is accurately described by an energy-dependent non-linear integral equation. However, the non-linear model is not invertible with a computationally efficient solution, and is often approximated as a linear integral model in the form of the Radon transform. Such approximation basically ignores energy-dependent information and would generate beam hardening artifacts. Dual-energy CT (DECT) scans one object using two different x-ray energy spectra for the acquisition of two spectrally distinct projection datasets to improve imaging performance. Thus, DECT can reconstruct energy and material-selective images, realizing monochromatic imaging and material decomposition. Nevertheless, DECT would increase radiation dose, system complexity, and equipment cost relative to single-spectrum CT. In this paper, a machine-learning-based CT reconstruction method is proposed to perform monochromatic image reconstruction using a single-spectrum CT scanner. Specifically, a residual neural network (ResNet) model is adapted to map a CT image to a monochromatic counterpart at a pre-specified energy level. This ResNet is trained on clinical dual-energy data, showing an excellent convergence to a minimal loss. The trained network produces high-quality monochromatic images on testing data, with a relative error of less than 0.2%. This work has great potential in clinical DECT applications such as tissue characterization, beam hardening correction and proton therapy planning.