Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session A: Tuesday, August 12, 1:30 – 4:30 pm, de Brug & E‑Hall

Functional Inter-Subject Alignment Outperforms Anatomical Alignment on fMRI Data in Inter-Subject Information Transfer

Leo Michalke¹, Jochem W. Rieger¹; ¹Carl von Ossietzky Universität Oldenburg

Presenter: Leo Michalke

Group-level analysis in neuroscience requires precise alignment of neural data across participants. A common approach is to spatially align brains and use spatial smoothing to enhance inter-individual overlap. However, this purely anatomical approach makes strong assumptions about functional-anatomical coupling that are likely violated due to substantial inter-individual variability in functional neuroanatomical organization. In this work, we compare multiple methods which aim to find a common representation of functional magnetic resonance imaging (fMRI) data across participants. The data was recorded from 30 participants listening to naturalistic auditory stimuli (Forrest Gump audio movie). We compare the standard anatomical approach (MNI space combined with spatial smoothing) and three inter-subject alignment methods (multiset canonical correlation analysis (MCCA), Kettenring 1971; group ICA, Calhoun 2009; Hyperalignment, Xu 2012) which seek to find a functional alignment by maximizing similarity of activation time-series between subjects in a latent space under different constraints. In order to evaluate the inter-subject information transfer of the different alignment methods, we designed a classification task based on decoding the occurrence of function versus content words in the audio movie. Inter-subject classifiers were trained on the aligned data from one set of subjects and tested on a held-out subject in a leave-one-subject-out fashion. The results show that functional inter- subject alignment methods (accuracy: MCCA, 0.638; ICA, 0.637; Hyperalignment, 0.611; chance level 0.5) greatly outperform the standard anatomical alignment method (MNI space, 0.508). This indicates that the important features for across subject generalization lie within the latent functional spaces, while anatomical-functional representations can be idiosyncratic. Our work demonstrates that functional inter-subject alignment has the potential to improve the generalizability of data representations when combining data of different subjects.

Topic Area: Methods & Computational Tools

Extended Abstract: Full Text PDF