Assessing variations in 3D image quality in chest CT across sites and scanners

BACKGROUND: Large CT image databases are critical to the development and validation of protocol harmonization strategies and AI models. However, image quality (IQ) can vary significantly due to differences in technology and site practices, even for scanners from the same manufacturer, which can have considerable effects on harmonization efforts and AI algorithm development. Quantifying the 3D IQ variability represented in these databases is essential to characterize protocol heterogeneity, the training data distribution, and to assess test data overlap. PURPOSE: To quantify and characterize sources of 3D IQ variability in chest CT across scanner models and imaging sites. METHODS: levels of 2.1 (low dose "LD"), 4.2 (MD), and 6.3 mGy (HD), and provide four reconstructions, including standard 'Std' and thin 'Thn' slice thicknesses for sharp "Sh" and soft "So" kernels. Variance component analysis (VCA) was performed on the resulting phantom series to quantify sources of variability in IQ partitioned into inter-scanner model, inter-site, and residual (i.e., unexplained) components. RESULTS: for So-Thn-HD (54%). CONCLUSION: This multisite phantom study demonstrates that substantial, systematic variability in 3D IQ exists across scanners, sites, metrics, dose levels, and reconstruction settings despite use of a routine non-contrast chest CT protocol. Variability was primarily driven by scanner model differences, with additional contributions from inter-site differences and residual variance for select metrics and protocols. Regardless of controlled imaging conditions that limit variability from patient anatomy, size-dependent dose effects, and tube current modulation, substantial variability in 3D IQ persists. The size-specific, relative trends observed in this study emphasize the importance of accounting for scanner model, site, and protocol-specific heterogeneity when using large CT image databases.