Purchase this article with an account.
Yichen Li, Rebecca Saxe, Stefano Anzellotti; Intersubject multivariate connectivity reveals optimal denoising strategies for visual category-specific regions. Journal of Vision 2019;19(10):257. doi: https://doi.org/10.1167/19.10.257.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Recognizing face and scene images recruits distinct networks of brain regions. Investigating how information is processed and transformed from region to region within these networks is a critical challenge for visual neuroscience. Recent work has introduced techniques that move towards this direction by studying the multivariate statistical dependence between patterns of response (Coutanche and Thompson-Schill 2013, Anzellotti et al. 2017, Anzellotti and Coutanche 2018). As researchers develop increasingly sophisticated tools to model statistical dependence between the fMRI signal there is a risk that models may increasingly capture artifactual relationships between regions. Choosing optimal denoising methods is a crucial step to maximize the accuracy of connectivity models. A common approach to compare denoising methods uses simulated fMRI data, but it is unknown to what extent conclusions drawn using simulated data generalize to real data. To overcome this limitation, we introduce intersubject multivariate pattern dependence (iMVPD) which computes the statistical dependence between a brain region in one participant, and another brain region in a different participant. IMVPD is multivariate, it trains and tests models on independent partitions of the real fMRI data, and it generates predictions that are both between subjects and between regions. Since whole-brain sources of noise are more strongly correlated within subject than between, we can use the difference between standard MVPD and iMVPD as a ‘discrepancy metric’ to evaluate denoising techniques, where more effective techniques should yield smaller differences. As a sanity check, the ‘discrepancy metric’ is the greatest with no denoising. Furthermore, a combination of CompCorr and removal of the global signal optimizes denoising in face- and scene-selective regions (among all denoising options tested on selected regions). In future work, iMVPD can be used for applications like studying individual differences and analyzing types of data where only one region is measured in each participant (i.e. electrophysiology).
This PDF is available to Subscribers Only