Benjamin Cheng Yin¹, Ana Clara Villegas Botero¹, Max Lobel¹, Mingzhou Ding¹; ¹University of Florida

Presently, to identify brain regions engaged in visual emotional processing using fMRI data, one of three approaches is typically applied: univariate analysis based on general linear model (GLM), within-subject multivariate decoding analysis (wMVPA) where the classifiers are trained and tested within each subject, or across-subject multivariate decoding analysis (aMVPA) where the classifiers are trained on N-1 subjects and tested on the remaining subject. To understand the similarities and differences among these methods, we analyzed fMRI-BOLD responses from N=20 participants viewing pleasant, neutral, and unpleasant natural scene images drawn from the International Affective Picture System (IAPS). For univariate analysis, two contrast maps were generated: pleasant-neutral and unpleasant-neutral. For the two MVPA methods, a whole-brain searchlight analysis was carried out to generate two decoding accuracy maps: pleasant-neutral and unpleasant-neutral. At voxel level, the strongest correlations were between the two MVPA methods (pleasant-neutral r = 0.63; unpleasant-neutral r = 0.66), followed by aMVPA and univariate analysis (r = 0.39; r = 0.47), with the weakest correlations found between wMVPA and univariate analysis (r = 0.34; r = 0.37). At the brain region level, across all three methods, the most consistent effects were observed in posterior visual and lateral parietal cortices. wMVPA and aMVPA further converged on fusiform cortex, inferior parietal lobule, supramarginal gyrus, superior parietal lobule, precuneus, and cerebellar lobule VI. Conversely, univariate GLM analysis revealed significant effects in bilateral amygdala, hippocampus, posterior cingulate cortex, posterior orbitofrontal cortex, mid temporal cortex, bilateral Heschl’s gyrus, superior temporal cortex, mid temporal cortex, thalamus, and bilateral precuneus. These results suggest that (1) convergence between univariate and multivariate methods is low, (2) the two MVPA methods, while showing more convergence, can still differ significantly, and (3) applying all three methods together will yield the most comprehensive understanding of the brain regions engaged in visual emotion processing.

Acknowledgements: NIH grants MH125615 and MH112558 and NSF grant BCS2318984