Figures and data
Single-contrast vs multi-task functional localization
(a) Single-contrast localizer. The activity of each voxel (dot) is characterized along a single dimension which compares its activity for Task a relative to a control task (Task b). All voxels exceeding a certain activity threshold (gray line) are selected as the desired ROI. (b) Multi-task localizer. The activity of each voxel is characterized by its response to multiple tasks relative to the mean across all tasks. Voxels located at the origin of the coordinate system respond equally to Task a, b and c. Different regions are defined by their response profile across all tasks (red, green and blue arrows). Voxels are assigned to the region with the most similar profile, independent of the length of their activity profile (distance from origin).
Simulated comparison of localization methods
(a) Distribution of subject-specific fSNRs estimated from the MDTB dataset. The line indicates the best-fitting gamma distribution. (b) Estimated ROI size as a function of fSNR for the three localizers. Each dot represents one simulated subject. Regression lines are shown for each of the localizers. (c) Estimated ROI size as a function of true ROI size for the three localizers. Each dot represents one simulated subject. Regression lines are shown for each of the localizers. (d) Average localization accuracy across localizer methods. Accuracy measured as the Dice coefficient between the estimated ROI and true region across simulated individuals. Error bars indicate across-subjects standard deviation of the differences between localizer methods.
Empirical validation in cerebellar language localization
(a) Percentage of individuals in which each cerebellar voxel was assigned to the language ROI using the contrast-fixed localizer. (b) Percentage of individuals in which each cerebellar voxel was assigned to the language ROI using a multi-task localizer. (c) Average inter-subject Dice coefficient for each localizer. Error bars indicate standard error of the mean across subjects. (d) Selectivity and specificity of single-contrast and multi-task language localizers in the cerebellum. Mean activation across individuals for a target and a control contrast inside and outside the defined language ROI (constrained to the right cerebellar hemisphere). Error bars indicate standard error of the mean across subjects.
Optimal battery selection strategies across battery sizes.
(a) Parcellation simulations. Dice overlap between estimated and true parcellations. Shaded area indicates standard deviation across 100 simulations with different task libraries. (b) Connectivity modeling simulations. Correlation between true connectivity weights and estimated connectivity weights. Shaded area indicates standard deviation across 100 simulations with different task libraries. (c) Parcellation of the neocortex in fMRI. Cosine similarity between true test data and predicted test data, averaged across subjects. Shaded area indicates standard error across subjects (N=24) of the differences between selection strategies in the MDTB dataset. (d) Neocortex-cerebellum connectivity modeling in fMRI. Cosine similarity between true cerebellar test data and predicted cerebellar test data, averaged across subjects. Shaded area indicates standard error across subjects of the differences between selection strategies in the MDTB dataset.
Experimental design of multi-task batteries.
(a) Average covariance of neocortical activity patterns across two imaging runs for the working memory and the language sessions of the HCP task dataset. (b) Proportion of the estimated baseline noise variance 
Temporal autocorrelation and task carry-over effects in interspersed designs.
(a) Mean covariance of task-evoked activity patterns in the MDTB dataset as a function of the difference in task positions (lag) within each imaging run. Error bars indicate standard error of the mean across subjects. (b) Mean covariance of the task-evoked activity patterns when a task is preceded by a different task vs the same task, shown for both across subjects (group) and within subject (individual) estimates. Error bars indicate standard error of the mean across subjects.
Group vs. individual task-by-task covariance matrices.
(a) Covariance matrix using an individual-level cross-validated approach. (b) Covariance matrix using group-averaged data without cross-validation. This library can be directly used in the MultiTaskBattery toolbox to select optimal task batteries for a brain structure of interest. Researchers can specify the target brain structure and the number of tasks, and the toolbox will apply the Minimal Collinearity criterion described in this paper to recommend tasks suitable for brain mapping studies.
Example optimal task batteries by region and battery size for an 8-minute fMRI scan.
Optimal battery selection strategies for brain parcellations of different brain structures.
(a) Prefrontal cortex. Cosine similarity between predicted and measured test data, averaged across subjects. Shaded area indicates standard error across subjects of the differences between selection strategies in the MDTB dataset. Regions are defined using 44 parcels of the Glasser atlas (Glasser et al., 2016; Donahue et al., 2018). (b) Cerebellum. Regions defined using the NettekovenSym32 atlas (Nettekoven et al., 2024).
Grouped vs Interspersed designs after accounting for instruction periods.
Predicted standard deviation of task-rest, and task-task contrasts for grouped (blue) and interspersed (orange) designs as a function of the proportion of each imaging run dedicated to rest. Lower panel shows the proportion of noise variance that is due to estimation of resting baseline. For the vertical line, the length of rest is equivalent to each single task condition.
