Rationale: Over the past 40 years, the neuroimaging field has successfully used magnetic resonance imaging to study brain structure and function. Precisely how researchers define structure is of paramount importance for creating models to predict function. Brain region definitions are especially important when using structural networks generated from diffusion tensor imaging (DTI) tractography to predict functional networks generated from intracranial recordings of neural electrical activity (Shah et al. 2019). In this study, we compare different standard brain atlases with atlases composed of randomly generated brain regions and show how atlas choice, parcellation size, and regional shape affect the statistical power to predict neural activity in other regions of the brain. Methods: Six medically refractory epilepsy patients underwent High Angular Diffusion Imaging (HARDI) and were subsequently implanted with Stereoelectroencephalography (Figure 1A.) We computed functional connectivity between electrodes using broadband cross-correlation in one-second intervals during 20 recorded seizures and corresponding interictal, preictal, and postictal periods (Figure 1B). We computed structural connectivity from streamline counts in 12 commonly used neuroimaging atlases and 12 randomly generated whole-brain atlases of different parcellation sizes consistent with past studies (30 permutations each) (Zalesky et al. 2010). We computed the Spearman Rank Correlation between the structural and functional connectivity matrices over the course of each one-second interval for each of the 372 total atlases (Figure 1C). Finally, we examined how the structure-function correlation (SFC) changed as a function of atlas choice and parcellation size to determine an effective atlas size for predicting SFC using bootstrap resampling methods (Figure 2).
Results: The most effective atlas parcellation size to predict functional connectivity from structural connectivity is 2.5 cm3 [2.4 - 2.6 cm3 95% CI] (Figure 2A). In an epilepsy context, the optimal atlas parcellation size is 16.0 cm3 [14.8 - 17.3 cm3 95% CI] to detect a change in SFC as a seizure develops (Figure 2B). Comparing the range of all standard neuroimaging atlases used in this study, it would take only an average of 8 recorded seizures for the AAL atlas to detect a significant change in SFC (p < 0.05 after Bonferroni Correction using paired test and a null hypothesis of no difference between preictal and ictal periods) compared to 14 recorded seizures for the Schaeffer 100 atlas at 80% power. Conclusions: We determined the atlas parcellation scale to build suitable structural models of the brain to predict its function. We also identified the common neuroimaging atlases for effective statistical power in detecting a change in SFC. In epilepsy patients, predicting the change in SFC using the appropriate atlas parcellation scale will allow epileptologists to better understand specific structural connections implicated in functional changes during a seizure.