[Paper Review] A deep learning approach for virtual monochromatic spectral CT imaging with a standard single energy CT scanner

Wei Zhao, Tianling Lyu+2 more|arXiv (Cornell University)|May 20, 2020

Advanced X-ray and CT Imaging38 references4 citations

TL;DR

This paper proposes a deep learning (DL) method to generate high-quality virtual monochromatic CT (VMCT) images from standard single-energy CT (SECT) scans, bypassing the need for expensive dual-energy CT (DECT) scanners. By training a residual network to predict high-energy CT images from low-energy SECT scans, the approach achieves up to 68% noise reduction and 55% higher iodine contrast-to-noise ratio (CNR) compared to conventional DECT-based VMCT, with optimal CNR at 40 keV.

ABSTRACT

Purpose/Objectives: To develop and assess a strategy of using deep learning (DL) to generate virtual monochromatic CT (VMCT) images from a single-energy CT (SECT) scan. Materials/Methods: The proposed data-driven VMCT imaging consists of two steps: (i) using a supervised DL model trained with a large number of 100 kV and 140 kV dual-energy CT (DECT) image pairs to produce the corresponding high-energy CT image from a low-energy image; and (ii) reconstructing VMCT images with energy ranging from 40 to 150 keV. To evaluate the performance of the method, we retrospectively studied 6,767 abdominal DECT images. The VMCT images reconstructed using both DL-derived DECT (DL-DECT) images and the images from DECT scanner were compared quantitatively. Paired-sample t-tests were used for statistical analysis to show the consistency and precision of calculated HU values. Results: Excellent agreement was found between the DL-DECT and the ground truth DECT images (p values ranged from 0.50 to 0.95). Noise reduction up to 68% (from 163 HU to 51 HU) was achieved for DL-based VMCT imaging as compared to that obtained by using the standard DECT. For the DL-based VMCT, the maximum iodine contrast-to-noise ratio (CNR) for each patient (ranging from 15.1 to 16.6) was achieved at 40 keV. In addition to the enormous benefit of VMCT acquisition with merely a SECT image, an improvement of CNR as high as 55% (from 10.7 to 16.6) was attained with the proposed approach. Conclusions: This study demonstrates that high-quality VMCT images can be obtained with only a SECT scan.

Motivation & Objective

• To develop a method for generating virtual monochromatic CT (VMCT) images using only a standard single-energy CT (SECT) scanner, eliminating the need for high-cost dual-energy CT (DECT) scanners.
• To improve image quality of VMCT by reducing noise and enhancing contrast-to-noise ratio (CNR), particularly at low energies where iodine contrast is highest.
• To enable widespread clinical adoption of VMCT by making it accessible on existing SECT hardware through deep learning-based image synthesis.
• To validate the accuracy and consistency of Hounsfield unit (HU) values in DL-derived VMCT images compared to ground truth DECT.

Proposed method

• A supervised deep learning model is trained on paired 100 kV and 140 kV dual-energy CT (DECT) image pairs to predict the high-energy 140 kV image from a low-energy 100 kV SECT image.
• The model uses a dual-energy residual mapping network that outputs the residual image (difference between predicted and input low-energy image), which is added to the input to reconstruct the high-energy image.
• A fully convolutional network (FCN) is applied to denoise the input 100 kV SECT image before feeding it into the residual network to improve prediction fidelity.
• Virtual monochromatic images are reconstructed across 40–150 keV using the DL-predicted DECT images, enabling energy-specific imaging with optimized contrast.
• The method leverages the strong noise correlation between low- and high-energy images, preserving noise texture and reducing overall noise in VMCT reconstructions.
• Statistical validation via paired-sample t-tests confirms consistency in HU values between DL-derived and ground truth DECT images.

Experimental results

Research questions

• RQ1 Can deep learning accurately predict high-energy CT images from low-energy SECT scans to enable virtual monochromatic CT without a DECT scanner?
• RQ2 Does the DL-based VMCT approach achieve superior image quality in terms of noise reduction and contrast-to-noise ratio (CNR) compared to conventional DECT-based VMCT?
• RQ3 At what energy level is the maximum iodine CNR achieved in DL-based VMCT, and how does it compare to conventional DECT-based VMCT?
• RQ4 How consistent are the Hounsfield unit (HU) values in DL-derived VMCT images relative to ground truth DECT images across different patient anatomies?
• RQ5 Can the DL-predicted DECT images be used effectively for other DECT applications such as material decomposition or virtual non-contrast imaging?

Key findings

• The DL-based method achieved excellent agreement with ground truth DECT images, with p-values ranging from 0.50 to 0.95 in paired-sample t-tests, indicating high consistency in HU values.
• Noise was reduced by up to 68% (from 163 HU to 51 HU) in DL-based VMCT compared to standard DECT-based VMCT, due to preserved noise correlation between low- and high-energy images.
• The maximum iodine CNR across all patients was 16.6, achieved at 40 keV, which is 55% higher than the conventional method’s maximum of 10.7.
• The optimal energy for maximum CNR shifted from 80–85 keV in conventional VMCT to 40 keV in the DL-based approach, enhancing visibility of low-contrast lesions.
• The DL-predicted high-energy images showed non-linear mapping behavior, with HU values differing between anatomical structures (e.g., kidney vs. bone marrow) even when input HU values were similar.
• The method enables high-quality VMCT imaging on standard SECT scanners, with potential for broader clinical use in applications like lesion detection, metal artifact reduction, and virtual non-contrast imaging.