From Intrinsic Topography to Cognitive Prediction: Benchmarking Approaches to Personalized Cortical Parcellation

Abstract

Personalized functional magnetic resonance imaging (fMRI) adapts cortical parcellations to individual functional anatomy. Various methods have been proposed that differ in how they combine group-level information, spatial constraints, and individual functional data. Using the Human Connectome Project Young Adult dataset (N = 947), we compared five personalized parcellation methods and a standard group-level atlas. For each method, we calculated functional homogeneity, test-retest reliability, parcel contiguity, and inter-individual Dice overlap. Personalized parcellation methods generally yielded higher functional homogeneity and lower inter-individual Dice than the group atlas, while test-retest reliability of functional connectivity was high across methods. Parcel contiguity differed across methods: some approaches produced almost entirely contiguous parcels, whereas others frequently produced spatially disconnected components, indicating that the detection of candidate ectopic variants depends partly on the parcellation strategy. We also examined whether connectivity and topographic features were associated with and predicted general intelligence (g). Using structural equation modeling, intra-individual parcel size variability and inter-individual parcel size similarity derived from the hierarchical Bayesian method (gMSHBM) were positively associated with g, whereas no corresponding association survived correction for the other methods. Using kernel ridge regression, functional connectivity significantly predicted g, with the group atlas yielding the highest predictive accuracy, consistent with a potential advantage of cross-individual spatial correspondence for edge-wise prediction. Finally, we assessed robustness across four preprocessing pipelines. Although gMSHBM parcellations were not re-estimated, relative between-method patterns in several spatial characteristics were largely preserved. In contrast, brain-behavior associations involving time-series-derived features varied substantially with preprocessing, whereas parcel-size associations were more stable. We conclude that personalized parcellation approaches are associated with distinct trade-offs in spatial correspondence, topographic variability, and downstream predictive utility. The choice of parcellation strategy therefore affects both the properties of individualized maps and the brain-behavior information derived from them.