Abstract Rationale Low-dose chest computed tomography (LDCT) enables early lung cancer detection and quantification of related imaging biomarkers. The National Lung Screening Trial (NLST) is a multicenter longitudinal study of LDCT scans. However, heterogeneous acquisition parameters, specifically, the reconstruction kernel, impact the texture of the underlying anatomy: hard kernels sharpen the image but introduce pixel noise while soft kernels smoothen the image with reduced pixel noise. The pixel noise results in variable quantitative measures in percent emphysema quantification, making it difficult to obtain accurate and comparable measures. Therefore, it is necessary to standardize the images to a common reference kernel. Methods We extended our previously developed anatomy-guided multipath cycleGAN framework to a larger NLST cohort and additional reconstruction kernels. The model was applied to nine reconstruction kernels (Siemens B50f, Siemens B30f, Siemens B80f, GE BONE, GE STANDARD, GE LUNG, Philips B, Philips C and Philips D) to enable multi-domain kernel harmonization though a shared latent space. Images were harmonized to the style of commonly used Siemens kernels for emphysema quantification: B30f (soft kernel) and B50f (hard kernel). We evaluated harmonization on 22616 subjects with 36607 scans at timepoint T0 of the NLST. Differences in emphysema before and after harmonization were quantified using Cohen’s d with 95% bootstrap confidence intervals, accounting for unequal sample sizes across kernels. Results Prior to harmonization, emphysema scores demonstrated substantial kernel dependent variability (Figure 1), with medium to large effects sizes for all kernels (Cohen’s d 0.8; all 95% CIs excluding zero). Harmonization to Siemens B30f reduces kernel variability, yielding smaller effect sizes (Cohen’s d = 0.3) with confidence intervals showing overlap in majority of the kernels, however, the harmonized Philips D kernel exhibited residual bias. Harmonization to Siemens B50f achieved consistency across all kernels with uniformly small effect sizes, suggesting that kernel differences were mitigated post harmonization in the NLST cohort. Conclusions By extending our prior anatomy-guided multipath cycleGAN to a larger cohort in the NLST, we demonstrate that harmonization through a shared latent space consistently mitigates kernel variability in emphysema quantification across nine reconstruction kernels from three different manufacturers. Our expanded evaluation confirms that harmonized emphysema measures remain stable across previously studied and newly introduced kernels, with a residual bias observed for the Philips D kernel harmonized to the style of B30f. These findings establish a foundation for extending this framework for evaluating additional CT biomarkers. This abstract is funded by: National Cancer Institute (NCI)
Krishnan et al. (Fri,) studied this question.