Robustness, spatial detail, and pitfalls of fixed ICA dimensionality in resting-state fMRI networks at 1.5, 3, and 7 T

Resting-state fMRI functional connectivity analysis is usually performed with seed-based methods that strongly rely on user-dependent definitions of regions of interest. Data-driven methods like independent component analysis (ICA) can mitigate this need. However, the number of components that should be expected in an fMRI acquisition, which determines the model order of the ICA, is not defined, and it is not uniformly chosen across studies. This variability is further complicated by the dependence of component number on field strength, with higher field strengths typically yielding more detectable components. Therefore, relying on a predetermined number may influence the results. Here, we compare functional maps obtained through ICA analysis at different magnetic field strengths and at various levels of spatial detail. Our results confirm the presence of the most frequently reported resting-state networks across field strengths and demonstrate that higher magnetic field strength enables more robust detection of functional networks with greater spatial detail. We also show that: (1) fixing the number of components, although improving interpretability of group results, may provide an incomplete picture of brain function; (2) a greater number of components is consistently identified at higher field strength, suggesting that the model order should be adapted according to both field strength and spatial detail.