Few studies explored how magnetic field strength affects the reliability of network identification, especially with data-driven methods. Here we acquired resting-state fMRI at multiple field strengths on the same group of human participants (11 subjects at 1.5T; 10 at 3T; 9 at 7T; native isotropic resolutions of 3.0mm, 2.2mm and 1.8mm respectively) to compare results across field strengths. We used Independent Component Analysis (ICA) and varied smoothing levels and number of components. After pre-processing, 4, 6, 8, or 12mm smoothing, and regression of head motion, white matter and cerebrospinal fluid, group-ICA (gICA) was performed using MELODIC to identify resting-state networks (RSNs). Fixed (20) and automatic component dimensionality were compared using spatial correlation. The number of matching components (i.e., with Pearson’s correlation >25%) was taken as an indicator of the robustness at each spatial detail. Results showed that, at 7T, components were consistently reproduced across smoothing levels, but correlations were lower without fixed dimensionality, indicating component splitting. Matching components at 8mm and 12mm were consistent across all field strengths, but only 7T maintained similar matching at 4mm. Thalamic and cerebellar components were absent at 7T unless dimensionality was unfixed. The findings suggest that higher field strengths increase spatial detail. Automatic calculation of the number of components caused consistent component splitting, indicating that fixing the number of components may limit a comprehensive representation of function.

Acknowledgement

This study was partially supported by the Italian Ministry of Health under the grant “RC 2024” to IRCCS Fondazione Stella Maris.