A Probabilistic Functional Atlas of Human Occipito-Temporal Visual Cortex

Abstract

Human visual cortex contains many retinotopic and category-specific regions. These brain regions have been the focus of a large body of functional magnetic resonance imaging research, significantly expanding our understanding of visual processing. As studying these regions requires accurate localization of their cortical location, researchers perform functional localizer scans to identify these regions in each individual. However, it is not always possible to conduct these localizer scans. Here, we developed and validated a functional region of interest (ROI) atlas of early visual and category-selective regions in human ventral and lateral occipito-temporal cortex. Results show that for the majority of functionally defined ROIs, cortex-based alignment results in lower between-subject variability compared to nonlinear volumetric alignment. Furthermore, we demonstrate that 1) the atlas accurately predicts the location of an independent dataset of ventral temporal cortex ROIs and other atlases of place selectivity, motion selectivity, and retinotopy. Next, 2) we show that the majority of voxel within our atlas is responding mostly to the labeled category in a left-out subject cross-validation, demonstrating the utility of this atlas. The functional atlas is publicly available (download.brainvoyager.com/data/visfAtlas.zip) and can help identify the location of these regions in healthy subjects as well as populations (e.g., blind people, infants) in which functional localizers cannot be run.

Keywords: cortex-based alignment; human brain atlas; object recognition; retinotopy; visual cortex.

Figure 1

Example probabilistic group maps in the left hemisphere after two brain alignments. (A) Three example ROIs are displayed where the most left column, v1d, shows an early visual cortex map, and the middle and right columns display two higher-order visual category-selective regions in VTC, Cos-places, and mFus-faces. Probability values range from 0 to 1 where 0 indicates no subject at a given vertex and 1 that all subjects in the probabilistic maps shared the given vertex. mFus-faces reveals less consistency as shown by a lower percentage of yellow-colored vertices. Bottom inset displays zoomed in location of the main figure. (B) Same ROIs as in A but after NVA. Bottom inset for CoS-places and mFus-faces indicates the location of the axial slice in the volume.

Figure 2

LOOCV predictability analysis using the Dice coefficient (DSC) for retinotopic regions (A) and category-selective regions (B). x-axis: threshold of the probability map generated using N-1 subjects; y-axis: DSC. A DSC value of 1 indicates perfect overlap between the N-1 group map and the left-out subject; 0 indicates no overlap. Blue lines: DSC after CBA; red lines: DSC after NVA. Dark colors/top rows correspond to left hemisphere data, light colors/bottom rows to right hemisphere data. Red: face-selective ROIs; green: body-selective ROIs; yellow: character-selective ROIs; gray: motion-selective ROI; error bars: standard error (SE) across the N-fold cross-validation.

Figure 3

fROI size across occipito-temporal cortex. Average ROI size in surface space separately for the left hemisphere (LH, light gray) and right hemisphere (RH, dark gray). Error bars: SE across subjects. Regions of x-axis are organized by category.

Figure 4

Reproducibility of category-selectivity responses. For the two cortical expanses that contain the category-selective regions of the visfAtlas, VTC and LOTC, the reproducibility of category responses was computed across the t-contrast maps of single runs for each respective category (see Materials and Methods for details). Dark gray bars represent the Dice coefficient results based on t-contrast maps that were thresholded with t > 2.2, which equals P < 0.01, while light gray bars were based on t-contrast maps that were thresholded at t > 0. Error bars represent SEs across subjects.

Figure 5

MPM of occipito-temporal cortex fROIs. (A) visfAtlas in surface space after CBA. Each color displays a unique fROI group map thresholded at 0.2 of all subjects in which the given fROI could be identified. (B) Volume atlas using the same color coding as in surface space. Inset between coronal and axial view displays the slice location for coronal and axial slices, respectively. LH: left hemisphere; RH: right hemisphere. (C) A new group average brain (BVaverage) published in BrainVoyager, based on 20 adults. This average brain can be used for future studies as a common reference brain.

Figure 6

Correspondence between the visfAtlas and five VTC probabilistic maps from independent data. (A) We compared visfAtlas MPM fROIs (white outlines) in VTC with probabilistic maps (colored regions) of six functional regions from an independent dataset that used a similar localizer, which has been published previously (Stigliani et al. 2015; Weiner et al. 2017). Top: left hemisphere; bottom: right hemisphere. (B) Average Dice coefficient between fROIs of the individual subjects from Stigliani and Weiner and colleagues and the MPMs of our visfAtlas fROIs. Error bars: SEs across subjects. LH: left hemisphere; RH: right hemisphere.

Figure 7

Proportion of voxels that show maximum responsivity in left-out subjects are largely their own category. Using our volumetric atlas data, we generated a cross-validated estimate of voxel maximum responsivity in a left-out subject. N-1 times, we generated a volumetric MPM and calculated the proportion of voxel that were maximally responsive for the ROI’s category, e.g., face response voxel in mFus-faces. This gives an estimate for the expected specificity of the atlas. For each major category—faces, bodies, places, and characters—proportions of category responsivity are displayed with each region’s preferred category as the bottom bar of each stacked bar graph. Error bars: proportion own category selectivity across all left-out subjects.

Figure 8

Comparison of the visfAtlas to other probabilistic atlases. In A–D each red-yellow map is the probabilistic map of unthresholded individual regions of the visfAtlas ROI and the outline is the fROI of the relevant atlas; all images are show in the fsaverage brain. (A) Comparison of V1-V3 dorsal and ventral of the retinotopic atlas published by Wang et al. (2015) and our respective visfAtlas regions. Regions are presented on a medial-occipital view of the fsaverage group brain. (B) Comparison of V1–V3 dorsal and ventral to the anatomically estimated V1–V3 (Benson et al. 2012). (C) Comparison of motion-selective hMT+ published by Huang et al. (2019) to visfAtlas hMT+ probabilistic map. (D) Comparison of CoS-places published by Weiner et al. (2018) to the visfAtlas CoS-places map. (E) Dice coefficient between the visfAtlas fROI and the same fROI defined by other atlases. Error bars: SE across 19 visfAtlas subjects. LH: left hemisphere; RH: right hemisphere.