Cancer radiomic feature variations due to reconstruction kernel choice and integral tube current.

Abstract

Purpose: Radiological cancer imaging features, or radiomics features, can be derived to diagnose disease or predict treatment response. However, variability between vendors, scanners, protocols, and even reconstruction software versions is an obstacle to the clinical use of radiomics features. This study aimed to characterize the impact of kernel reconstruction differences and integral tube current settings on radiomic features extracted from computed tomography (CT) scans. Methods: Radiomic features were extracted from CT scans of a 3D-printed phantom with five imprinted tumors using the CERR software system, resulting in 282 features. Batch effects were assessed via principal component analysis (PCA) and correlation measures. Robustness was measured using the concordance correlation coefficient (CCC) and Pearson correlation coefficient. Statistical analysis was performed using R software. Results: PCA identified two clusters comprised of Standard, ASIRs, ASIRV, and soft kernels in one, and Lung and Bone Kernels in the other. Features displayed a gradient from ASIR10 to ASIR50 and ASIRV1 to ASIRV5 in terms of nearness to Standard Kernel feature values. Feature correlation matrices revealed little change in ASIRs, ASIRVs, and the Standard Kernel, but showed significant changes in Bone and Lung Kernel results. Combat-algorithm correction improved robustness, particularly in first-order statistic features, and mitigated batch effects due to the ASIRs and the standard kernel. Forty (40) out of 282 features were identified as robust. However, Combat-based correction performed poorly in harmonizing Bone and Lung reconstruction kernels. Conclusions: The robustness of means and median radiomic features across kernel reconstruction choices, in contrast to the lack of robustness in many other radiomic features, suggests that kernel reconstruction effects are not well-addressed by current harmonization methods.