Deducing Resting-State fMRI Seeds with Manifold Learning and Distance Clustering.

Motivation: Canonical approaches for automating seed-based correlation analysis (SCA) forgo personalization. Goal(s): Develop a personalized SCA algorithm that does not assume seed geometry or coordinate. Approach: For 79 patients with presurgical resting-state fMRI (rs-fMRI), seeds in primary language areas were obtained by determining inter-voxel temporal similarity with PaCMAP and HDBSCAN (P-H). Per patient, binary seeds from 1000 P-H iterations were consolidated into probabilistic seeds for SCA. Maximum dice coefficients with task-based (tb-fMRI) localizations were compared to existing method. Results: SCA with probabilistic seeds identified by this new method had significantly higher dice coefficients with tb-fMRI language localizations compared to existing method. Impact: Personalized SCA with P-H method and probabilistic seeds improves the accuracy of detecting the rs-fMRI language network in brain tumor patients. This can facilitate clinical adoption of rs-fMRI for patients needing presurgical language localization but have limited tb-fMRI results.