Influence of computed tomography reconstruction algorithms on coronary artery calcium scores and reader agreement.

BACKGROUND: Coronary artery calcium scoring (CACS) is central to cardiovascular risk stratification. Differences between reconstruction algorithms may introduce inconsistencies in Agatston scores, resulting in reduced reliability and diminished clinical utility.

METHODS: 120 CACS scans were reconstructed using filtered back projection (FBP), iterative reconstruction (ASIR-V), and deep-learning-based image reconstruction (DLIR-H), and independently scored by two blinded readers. Intra-reader and inter-reader agreement were assessed categorically using Bayesian cumulative logit mixed-effects modeling and weighted kappa, and continuously using intraclass correlation coefficients (ICCs) of log-transformed scores. Reader confidence was compared using Friedman and Wilcoxon tests.

RESULTS: Categorical inter-reader agreement was highest for DLIR-H (agreement 94.2%, κ ​= ​0.994), followed by ASIR-V (90.0%, κ ​= ​0.984) and FBP (86.7%, κ ​= ​0.976). Inter-reader ICCs were significantly higher for DLIR-H (0.996) than both FBP (0.974, p ​= ​0.002) and ASIR-V (0.984, p ​< ​0.001). FBP and ASIR-V both exhibited a reader-dependent bias, with one reader assigning higher calcium scores (categorical: FBP OR 11.1 [3.7-36.0], ASIR-V OR 3.8 [1.3-12.1]; continuous: Δlog score +0.23 and ​+ ​0.11, respectively), a discrepancy not present with DLIR-H. Within-reader contrasts showed higher Agatston scores for FBP than DLIR-H for one reader (categorical: OR 4.03 [1.29-12.14], continuous: Δlog score +0.16). Mean reader confidence differed across algorithms and was greatest for DLIR-H.

CONCLUSIONS: DLIR-H, a deep-learning-based reconstruction algorithm, significantly enhances inter- and intra-reader reliability and reader confidence for CACS compared with FBP and ASIR-V, offering significantly improved reproducibility of CACS for cardiovascular risk assessment.