Artificial intelligence model-based iterative reconstruction for lung ultra-low-dose CT: image quality, ground-glass nodules detectability, and Lung-RADS evaluation

  1. Ying Chen
  2. Jing Deng
  3. Zhuo Zhu
  4. Xi Cai
  5. Jiashuai Li
  6. Wei Xu
  7. Chenglin He
  8. Qing Zhang
  9. Jianxin Cao
Read the full article See related articles

Listed in

This article is not in any list yet, why not save it to one of your lists.
Log in to save this article

Abstract

Objective: To assess the effect of artificial intelligence model-based iterative reconstruction (AIIR) on image quality of lung ultra-low-dose CT (ULDCT), as well as its influence on the detection and diagnostic classification of GGNs. Methods: Fifty-three patients diagnosed with GGNs underwent both lung standard-dose CT (SDCT) and ultra-low-dose CT (ULDCT) scans. SDCT hybrid iterative reconstruction (SDCT-HIR) images, ULDCT hybrid iterative reconstruction (ULDCT-HIR) images, and ULDCT-AIIR images were generated using a lung sharpness algorithm. Image noise measurements were performed. Image quality was independently scored by Radiologists 1 and 2. Separately, GGNsdetectability and Lung-RADS classification were independently evaluated by Radiologists 3 and 4. Results: ​Two patients were excluded due to respiratory motion artifacts and one due to raw data errors, resulting in 50 patients with 81 GGNs for final analysis. Compared to SDCT, ULDCT achieved a remarkable 93.7% reduction in radiation dose​(SDCT: 6.27 ± 0.98 mSv vs. ULDCT: 0.40 ± 0.13 mSv, P < 0.001). Moreover, ULDCT-AIIR images exhibited the lowest noise levels (P < 0.001). The image quality scores of ULDCT-AIIR were significantly superior to those of ULDCT-HIR (P < 0.001) and comparable to SDCT-HIR (P > 0.05). For GGNs detection, Radiologist 3 reported rates of 64.2% (ULDCT-HIR) vs. 95.1% (ULDCT-AIIR), while Radiologist 4 reported 67.9% (ULDCT-HIR) vs. 96.3% (ULDCT-AIIR). ULDCT-AIIR images demonstrated significantly higher detection rates than ULDCT-HIR images (P < 0.001). The consistency in the Lung-RADS classification was moderate between ULDCT-HIR and SDCT-HIR images (κ=0.343 and 0.411 for radiologist 3 and 4, respectively), but good between ULDCT-AIIR and SDCT-HIR images (κ=0.772 and 0.743 for radiologist 3 and 4, respectively). Conclusion: AIIR can significantly enhance image quality in ULDCT while maintaining excellent GGNs detection capability. Moreover, there is good consistency in Lung-RADS classification between ULDCT-AIIR and SDCT-HIR images.

Article activity feed

  1. Version published to 10.21203/rs.3.rs-7206057/v1 on Research Square