Repeatability analysis of ICA-based harmonization for multi-site MRI data using dual projection models

Integrating multi-site magnetic resonance imaging (MRI) datasets enhances statistical power and generalizability in neuroimaging research but introduces systematic variability, known as site effects, that can obscure true biological signals. Independent component analysis (ICA)-based harmonization methods, such as the dual projection ICA (ICA-DP) model, aim to mitigate these effects while preserving meaningful signal. However, the repeatability of ICA decompositions across different runs and parameter settings remains a critical challenge, affecting the stability and reliability of site effects removal. Here, we propose a novel evaluation framework for ICA repeatability that jointly assesses spatial components, their mixing coefficients, and component energy-a theoretically grounded but often overlooked parameter based on back-projection. Using both simulated and real multi-site MRI datasets, we demonstrate that incorporating component energy into repeatability metrics provides a more robust and theoretically grounded assessment of ICA stability. We further revise the ICA-DP harmonization scheme by removing site-associated components comprehensively, resulting in improved preservation of biologically relevant signals. Our results establish the importance of repeatability analysis and support the proposed framework as a reliable tool for ICA-based harmonization in multi-site studies.