Abstract
An unprecedented amount of existing functional Magnetic Resonance Imaging(fMRI) data provides a new opportunity to understand the relationship betweenfunctional fluctuation and human cognition/behavior using a data-drivenapproach. To that end, tremendous efforts have been made in machine learning topredict cognitive states from evolving volumetric images ofblood-oxygen-level-dependent (BOLD) signals. Due to the complex nature of brainfunction, however, the evaluation on learning performance and discoveries arenot often consistent across current state-of-the-arts (SOTA). By capitalizingon large-scale existing neuroimaging data (34,887 data samples from six publicdatabases), we seek to establish a well-founded empirical guideline fordesigning deep models for functional neuroimages by linking the methodologyunderpinning with knowledge from the neuroscience domain. Specifically, we putthe spotlight on (1) What is the current SOTA performance in cognitive taskrecognition and disease diagnosis using fMRI? (2) What are the limitations ofcurrent deep models? and (3) What is the general guideline for selecting thesuitable machine learning backbone for new neuroimaging applications? We haveconducted a comprehensive evaluation and statistical analysis, in varioussettings, to answer the above outstanding questions.