Induced T1 Degradation Selectively Destabilizes Surface-Mapped fMRI Representations

Surface-based representations are increasingly used in fMRI foundation models, and recent works show that surface inputs can outperform volume inputs on some downstream tasks. Surface inputs project BOLD timeseries onto a cortical mesh reconstructed from a structural T1-weighted (T1w) scan, coupling functional features to anatomical reconstruction quality even when the functional signal is unchanged. Here we systematically evaluate how stable surface and volume representations remain under anatomical degradation. We sequentially degrade the T1w image while holding functional BOLD timeseries fixed, and find that surface-mapped functional connectivity (FC) and secondary summary features diverge substantially more from their clean-anatomical baselines than volume-mapped equivalents. Foundation model embeddings inherit this same surface--volume asymmetry, though pretraining provides higher absolute representation stability. Downstream age and diagnosis prediction models trained on clean cohorts undergo severe performance degradation on surface-mapped test points, whereas volume-mapped test points remain stable. Surface-mapped foundation models and features are thus selectively fragile to anatomical scan quality: accuracy advantages measured on high-quality anatomy may not transfer under the variable scan conditions typical of real-world neuroimaging.