Comparison of image quality in carotid dual-energy computed tomography angiography at 55 keV virtual monoenergetic imaging using deep learning and adaptive iterative reconstruction algorithm

Xiaohan Liu, Chong Wang, Juan Long, Yang Wu, Zhongxiao Liu, Yu Meng, Chenzi Wang, Wenbei Xu, He Zhang, Aiyun Sun, Shuai Zhang, Kai Xu, Yankai MengJournal of Clinical Imaging Science 2025 · DOI

TL;DR

This study demonstrates that a high-strength deep learning image reconstruction (DLIR-H) algorithm significantly improves the image quality of 55 keV virtual monoenergetic imaging (VMI) in carotid dual-energy CT angiography (DE-CTA), outperforming traditional iterative reconstruction methods. This improvement is particularly pronounced in reducing noise and preserving fine anatomical details, especially in patients with a higher body mass index (BMI).

What Problem Does This Paper Address?

Evaluating carotid artery plaques is crucial for diagnosing and managing the risk of cerebrovascular events. While computed tomography angiography (CTA) is a standard imaging technique, dual-energy CTA (DE-CTA) offers enhanced plaque characterization. Low-energy virtual monoenergetic imaging (VMI) derived from DE-CTA data can improve vessel contrast and potentially reduce contrast agent and radiation doses. However, achieving high image quality at these lower energy levels, especially under "dual-low" conditions (reduced contrast and radiation), presents a challenge due to increased image noise. Traditional iterative reconstruction algorithms like ASIR-V can sometimes lead to over-smoothing, obscuring fine anatomical details. This paper investigates whether a newer deep learning-based image reconstruction (DLIR) algorithm, specifically a high-strength version (DLIR-H), can overcome these limitations and provide superior image quality in carotid DE-CTA compared to ASIR-V, particularly in challenging patient populations like those with higher BMIs.

Methods

The study employed a prospective design involving 48 patients undergoing DE-CTA. The core of the methodology lies in comparing image reconstruction techniques. Raw DE-CTA data was reconstructed using three algorithms:

  • ASIR-V 50%: A standard adaptive statistical iterative reconstruction algorithm at 50% strength.
  • DLIR-L: A low-strength deep learning image reconstruction algorithm.
  • DLIR-H: A high-strength deep learning image reconstruction algorithm.

All reconstructions were performed to generate 55 keV VMI datasets. Image quality was assessed using two approaches:

  • Objective Evaluation: This involved quantitative measurements of background noise (standard deviation of muscle tissue), signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR) at key anatomical locations (aortic arch, common carotid artery, carotid bifurcation, and internal carotid artery).
  • Subjective Evaluation: Two senior radiologists independently evaluated images for noise, artifacts, vascular continuity, vessel wall delineation, and diagnostic confidence using a 5-point scale.

Statistical analysis was performed using SPSS, with appropriate tests (ANOVA, Kruskal-Wallis) to determine significant differences between the groups.

Datasets

The study utilized data from 48 patients who underwent DE-CTA examinations at the authors' institution between December 2024 and January 2025. No new datasets were introduced or made publicly available.

Key Insights and Ideas

  • DLIR-H as a Noise Reducer: The high-strength DLIR algorithm demonstrated a remarkable ability to suppress image noise, significantly outperforming both ASIR-V 50% and DLIR-L. This is crucial for low-keV VMI where noise is a primary concern.
  • Preserving Fine Details: Beyond noise reduction, DLIR-H was found to preserve anatomical details, particularly the vascular wall margin, which is essential for accurate plaque assessment. This addresses a known limitation of some iterative reconstruction methods that can lead to over-smoothing.
  • Obesity as a Challenge, DLIR-H as a Solution: The study highlighted that DLIR-H's benefits were even more pronounced in patients with a BMI ≥24 kg/m². This is significant because obese patients often present with higher image noise due to increased tissue attenuation, making image interpretation more difficult. DLIR-H offers a potential solution to improve diagnostic accuracy in this population.
  • Dual-Low Advantage: The research was conducted under "dual-low" conditions (reduced contrast agent and radiation dose), demonstrating that DLIR-H can maintain superior image quality even when imaging parameters are optimized for patient safety.

Key Results

  • Objective Metrics: DLIR-H significantly reduced background noise and increased SNR and CNR compared to ASIR-V 50% and DLIR-L (P < 0.05). No significant differences were found between ASIR-V 50% and DLIR-L. CT attenuations were comparable across all groups.
  • Subjective Metrics: DLIR-H received significantly higher subjective scores for image noise reduction and vascular wall delineation compared to the other two methods (P < 0.05). The overall subjective image quality score for DLIR-H was also significantly higher.
  • BMI Subgroup: In patients with BMI ≥24 kg/m², the pattern of improved noise reduction and increased SNR/CNR with DLIR-H was consistent and statistically significant.
  • Radiation Dose: The average effective dose was 1.54 ± 0.09 mSv.

Limitations and Future Work

  • Effective Dose: While the study aimed for "dual-low" conditions, the authors acknowledge that the average effective dose (1.53 ± 0.09 mSv) could potentially be further reduced.
  • BMI Subgroup Specificity: Further investigation is needed to explore the potential of DLIR to optimize VMI images specifically in patients with BMIs ≥28 kg/m².
  • Single-Center Study: The generalizability of the findings requires validation through multi-center studies with larger sample sizes.
  • Comparison with Other Modalities: Future studies should compare DE-CTA with DLIR-H against established techniques like MRA and DSA to provide a more comprehensive understanding of its clinical applicability.

Summary updated

Abstract

At 55 keV VMI in carotid DE-CTA, DLIR-H significantly enhanced image quality, particularly by reducing noise and preserving fine anatomical structures. Its efficacy was especially notable in patients with BMI ≥24 kg/m<sup>2</sup>.
BibTeX
@article{liu2025comparison,
  author = {Xiaohan Liu and Chong Wang and Juan Long and Yang Wu and Zhongxiao Liu and Yu Meng and Chenzi Wang and Wenbei Xu and He Zhang and Aiyun Sun and Shuai Zhang and Kai Xu and Yankai Meng},
  title = {{Comparison of image quality in carotid dual-energy computed tomography angiography at 55 keV virtual monoenergetic imaging using deep learning and adaptive iterative reconstruction algorithm}},
  journal = {Journal of Clinical Imaging Science},
  year = {2025},
  doi = {10.25259/jcis\_109\_2025},
}
Figure