Optimizing Coronary CT Image Reconstruction With Deep Learning for Improved Quality: A Retrospective Study.

Abstract

Objective: To evaluate the impact of deep learning image reconstruction on image quality in CCTA compared with adaptive statistical iterative reconstruction (ASIR).

Materials and methods: CCTA data sets from 100 consecutive patients with suspected CAD were acquired with a Revolution Apex 256-row CT scanner, reconstructed with ASIR-V and DLIR-H, and subsequently analyzed. Image noise, SNR, and CNR in five regions of interest (25 mm) were calculated and t tested. The normality of quantitative variables was assessed using the Shapiro-Wilk test. For non-normally distributed data, the Mann-Whitney U test was applied. The concordance of HU values within specific ROIs was analyzed with Bland-Altman plots. Correlation between ASIR-V and DLIR-H was conducted using the Spearman rank correlation test.Subjective image analysis was conducted using a 5-point scale to evaluate noise level, vascular enhancement smoothness, artifact reduction, and diagnostic confidence. Intraclass correlation (ICC) was used to assess the reliability and consistency of subjective ratings among the reader.

Results: DLIR-H significantly reduced image noise across all ROIs (from 15% to 41%, all P<0.05), compared with ASIR-V. Mean SNR (ASIR-V vs. DLIR-H) were septum=4.3±1.7 versus 6.4±2.2; cavity of the left ventricle=24.3±8.3 versus 36.6±11.7; CNR: septum=8.2±2.5 versus 12.4±3.5; cavity of left ventricle= 28.2±9.1 versus 42.5±13.0. Spearman rank correlation ranged from 0.64 to 0.79 (P<0.05). Bland-Altman analysis showed good agreement between ASIR-V and DLIR-H, with no discernible patterns. Subjectively, DLIR-H significantly outperformed ASIR-V across all evaluated criteria (all P<0.05). ICC values indicated strong agreement among readers, demonstrating excellent reliability for most criteria and good reliability for vascular enhancement smoothness.

Conclusions: DLIR-H significantly improved CCTA image quality compared with ASIR-V, which contributes to a more accurate diagnosis in patients with suspected CAD.