A meta-analysis of fMRI decoding: Quantifying influences on human visual population codes

Abstract

Information in the human visual system is encoded in the activity of distributed populations of neurons, which in turn is reflected in functional magnetic resonance imaging (fMRI) data. Over the last fifteen years, activity patterns underlying a variety of perceptual features and objects have been decoded from the brains of participants in fMRI scans. Through a novel multi-study meta-analysis, we have analyzed and modeled relations between decoding strength in the visual ventral stream, and stimulus and methodological variables that differ across studies. We report findings that suggest: (i) several organizational principles of the ventral stream, including a gradient of pattern granulation and an increasing abstraction of neural representations as one proceeds anteriorly; (ii) how methodological choices affect decoding strength. The data also show that studies with stronger decoding performance tend to be reported in higher-impact journals, by authors with a higher h-index. As well as revealing principles of regional processing, our results and approach can help investigators select from the thousands of design and analysis options in an empirical manner, to optimize future studies of fMRI decoding.

Keywords: Decoding; MVPA; Meta-analysis; Objects; Patterns; Vision.

Figure 1. A flow-chart of the paper identification and selection process

See also Supplementary Table 1 for the complete list of included papers.

Figure 2. Modeled influence of voxel size on decoding strength

Predicted changes in decoding strength are shown for every 1 mm³ added to the size of acquired voxels. Positive values indicate greater decoding strength for each visual region.

Figure 3. Change in decoding strength with increasing exemplar counts in each class

The y-axis shows the regression coefficient from a model predicting classification performance based on the number of exemplars within classified categories. Positive values indicate a positive relation between classification performance and class exemplars. Negative values reflect inverse relations.

Figure 4. Influences on classification performance

Colors reflect standardized coefficients from a regression predicting each region’s decoding strength. Red indicates that higher values of each variable predict greater decoding. Blue indicates that lower values predict greater decoding. White reflects an absence of a relation. For binary variables, a value of 1 was assigned to the first option listed in the y-axis. Predictors with too few data-points for an ROI are striped-out. Asterisks indicate relations that generalize beyond the current sample of papers (p < 0.05).

Figure 5. Variables ordered by the strength and direction of their relationship with decoding performance

Variables at the top and bottom have the greatest influence (in a positive and negative direction respectively). For ease of viewing, V1, V2, V3 and V4 have been collapsed, as have LOC, FFA, PPA and VT regions.

Figure 6. Influences on pattern correlations

Colors reflect standardized coefficients from a regression predicting each region’s pattern discrimination based on correlations (i.e., 1−r). Red indicates that higher values predict greater pattern discriminability. Blue indicates that lower predictor values predict greater discriminability. White reflects an absence of a relation. For binary variables, a value of 1 was assigned to the first option listed in the y-axis. Predictors that have too few data-points, and that are irrelevant for correlations (e.g., classifier type) are striped-out. Asterisks indicate relations that generalize beyond the current sample of papers (p < 0.05).

Figure 7. The methodological space of the study set

A: The similarity matrix shows pairwise Pearson correlations between investigations based on their methodological decisions, grouped by investigated region. B: Mean correlation values extracted from the quadrants of panel A.