Impact of a deep-learning image reconstruction algorithm on image quality and detection of solid lung lesions

Abstract

Purpose

To compare the impact of a deep-learning image reconstruction algorithm (Precise Image) with an iterative reconstruction algorithm on image quality and detection of solid lung lesions in chest CT images.

Methods

All consecutive patients with at least one solid lung lesion diagnosed between December 2021 and February 2022 were retrospectively included. Images were reconstructed using Level 4 of the iterative reconstruction algorithm (i4) and the Standard/Smooth/Smoother levels of the deep-learning image reconstruction algorithm. Mean attenuation and standard deviation were measured by placing regions of interest in fat, muscle, trachea and solid lung lesions. The contrast-to-noise ratio between the lesion and the trachea was computed. Two radiologists assessed image noise and image smoothness, overall image quality and confidence diagnostic level using Likert scales. One radiologist also measured the large axis of the largest lesion. Statistical analyses was performed to compare outcomes obtained with the different algorithms.

Results

Thirty patients with a mean age of 70.0 ± 9.0 years (17 men) were included. The mean CTDI_vol was 6.3 ± 2.1 mGy. For all tissues, the contrast-to-noise ratio was similar for i4 and Standard level (p > 0.05) but increased significantly with other deep-learning image reconstruction levels compared to i4 (p < 0.05) and increased significantly from Standard to Smoother. Radiologists rated the image noise with a similar score between i4 and Standard level but decreased significantly between i4 and other deep-learning image reconstruction levels (p < 0.05) and from Standard to Smoother levels (p < 0.01). Overall image quality score were highest for the Smooth and Smoother levels.

Conclusion

Smooth and Smoother levels may now be used in clinical practice for chest CT acquisitions in solid lung lesion follow-up.