[Paper Review] Dual-energy CT imaging from single-energy CT data with material decomposition convolutional neural network

Tianling Lyu, Zhan Wu+4 more|arXiv (Cornell University)|May 30, 2020

Advanced X-ray and CT Imaging38 references4 citations

TL;DR

This paper proposes a deep learning method, FLESH-DECT, that reconstructs high-quality dual-energy CT (DECT) images from a single fully-sampled low-energy CT scan and a single high-energy projection view. By leveraging anatomical consistency and energy-domain correlation through a material decomposition convolutional neural network (MD-CNN), the method enables accurate DECT imaging with reduced radiation dose and simplified hardware requirements, achieving superior image quality comparable to standard DECT.

ABSTRACT

Dual-energy computed tomography (DECT) is of great significance for clinical practice due to its huge potential to provide material-specific information. However, DECT scanners are usually more expensive than standard single-energy CT (SECT) scanners and thus are less accessible to undeveloped regions. In this paper, we show that the energy-domain correlation and anatomical consistency between standard DECT images can be harnessed by a deep learning model to provide high-performance DECT imaging from fully-sampled low-energy data together with single-view high-energy data, which can be obtained by using a scout-view high-energy image. We demonstrate the feasibility of the approach with contrast-enhanced DECT scans from 5,753 slices of images of twenty-two patients and show its superior performance on DECT applications. The deep learning-based approach could be useful to further significantly reduce the radiation dose of current premium DECT scanners and has the potential to simplify the hardware of DECT imaging systems and to enable DECT imaging using standard SECT scanners.

Motivation & Objective

• To enable high-performance dual-energy CT (DECT) imaging using standard single-energy CT (SECT) scanners without hardware modifications.
• To reduce radiation dose in DECT by minimizing high-energy projection data to a single view.
• To exploit anatomical consistency and energy-domain correlation between low- and high-energy CT images via deep learning.
• To generate material-specific images such as virtual non-contrast and iodine maps without requiring dedicated DECT scanners.
• To simplify DECT implementation for clinical settings with limited access to premium DECT systems.

Proposed method

• A denoising convolutional neural network (CNN) is applied to the low-energy CT image to reduce noise before material decomposition.
• A material decomposition CNN (MD-CNN) is trained to estimate basis material decomposition maps from the low-energy image, enabling reconstruction of high-energy images at any energy level.
• The high-energy image is predicted by back-projecting a projection vector that is constrained to match the measured single-view high-energy projection, ensuring physical consistency.
• The method uses a projection-domain CNN that enforces consistency between the predicted high-energy image and the measured single high-energy projection.
• The approach is trained end-to-end on 5,753 slices from 22 patients using a 100 kV/Sn 140 kV scanning protocol.
• The framework is extendable to 3D geometries by adapting the network and projection operators to volumetric data.

Experimental results

Research questions

• RQ1 Can high-quality DECT images be reconstructed from a single fully-sampled low-energy CT and a single high-energy projection view?
• RQ2 Can deep learning effectively exploit energy-domain correlation and anatomical consistency to replace conventional DECT hardware?
• RQ3 To what extent can radiation dose be reduced while maintaining diagnostic image quality in DECT?
• RQ4 Can the proposed method generate accurate material-specific images such as virtual non-contrast and iodine maps without dedicated DECT scanners?
• RQ5 How robust is the method when applied to different scanning protocols or imaging protocols?

Key findings

• The FLESH-DECT method achieved high-fidelity DECT image reconstruction using only one high-energy projection view, significantly reducing radiation dose compared to standard DECT.
• The method produced virtual non-contrast and iodine quantification images with diagnostic quality comparable to standard DECT, as validated on clinical contrast-enhanced scans.
• The MD-CNN generated accurate material decomposition maps from 100 kV images, enabling synthesis of high-energy images at various energy levels (e.g., 120 kV, 150 kV) without retraining.
• The approach demonstrated superior noise suppression and artifact reduction compared to conventional reconstruction methods, even with sparse projection data.
• The method is compatible with existing SECT scanners, as the single high-energy projection can be acquired via a scout-view scan without hardware modifications.
• The model’s performance was sensitive to protocol mismatches; however, retraining on different protocols could resolve this limitation.