Influence of CT harmonization in longitudinal radiomics for NSCLC immunotherapy response prediction

This study investigates the variability of radiomic features in longitudinal CT scans from a multi-institutional NSCLC cohort and introduces a harmonization pipeline to improve predictive modeling of immunotherapy response. Baseline and follow-up CT scans from NSCLC patients treated with anti-PD-1/PD-L1 agents were analyzed, with two institutions combined for model training and internal testing, and a third institution serving as an external test set. To address variability from imaging parameters-such as scanner manufacturer, slice thickness, and noise-we applied image harmonization followed by feature harmonization using NestedComBat. This approach substantially reduced feature dependence on acquisition confounders (from 78.8% to 12.8%) and improved feature robustness across institutions. We further assessed the temporal consistency of radiomic features across longitudinal scans using the intraclass correlation coefficient (ICC). Image harmonization yielded the largest gains in stability (mean ΔICC = +0.021, p < 0.001), while the combined approach also enhanced longitudinal reliability (ΔICC = +0.014, p < 0.001). Finally, harmonization improved predictive performance for 6-month immunotherapy response, increasing the AUC from 0.695 to 0.768 in the internal test and from 0.692 to 0.802 in the external test. These results demonstrate that combining image- and feature-level harmonization enhances the robustness and temporal consistency of radiomic features, potentially supporting more reliable and generalizable predictive modeling across diverse datasets and clinical settings.