The Human Connectome Project 7 Tesla retinotopy dataset: Description and population receptive field analysis

1The central nervous system maps sensory inputs onto topographically organized representations. In the field of vision, researchers have successfully exploited functional magnetic resonance imaging (fMRI) to noninvasively measure visual field representations (“retinotopy”) in the living human brain (DeYoe et al.,; Engel, Glover, & Wandell,; Engel et al.,; Sereno et al., ). These efforts enable parcellation of visual cortex into distinct maps of the visual field, thereby laying the foundation for detailed investigations of the properties of visual cortex (Abdollahi et al.,; Benson, Butt, Brainard, & Aguirre,; Wang, Mruczek, Arcaro, & Kastner,; for a review, see Silver & Kastner,; Tootell, Dale, Sereno, & Malach,; Wandell, Dumoulin, & Brewer,; Wandell & Winawer, ).

2One of the datasets acquired by the Human Connectome Project (HCP; Ugurbil et al.,; Van Essen et al., ) was a 7T fMRI retinotopy experiment. This experiment, conducted in 181 healthy young adults, involved carefully designed stimuli and a substantial amount of fMRI data (30 min, 1,800 time points) acquired at high spatial and temporal resolution (1.6-mm isotropic voxels, 1-s sampling). Although retinotopy is routinely measured in small groups of subjects by individual laboratories in support of various research projects, to date there has not been a large publicly available set of retinotopic measurements.

3In this paper, we describe the design of the retinotopy experiment and demonstrate the analyses that we have performed on the fMRI data. We adopt a model-based analysis approach in which a computationally intensive nonlinear optimization is performed to determine parameters of a population receptive field (pRF) model (Dumoulin & Wandell,; Kay, Winawer, Mezer, & Wandell,; Wandell & Winawer, ). The results include estimates of pRF position (angle and eccentricity) and pRF size for each “grayordinate” (cortical surface vertex or subcortical voxel), and can be used to define retinotopic maps in the brain. We show that the HCP retinotopy data provide high-quality pRF results in many parts of occipital, temporal, parietal, and frontal cortex. We make freely available these pRF results, as well as associated stimuli and analysis code, at an Open Science Framework website (https://osf.io/bw9ec/). The pRF results are also accessible via the Brain Analysis Library of Spatial Maps and Atlases (BALSA) database (https://balsa.wustl.edu/study/show/9Zkk; Van Essen et al., ), downloadable as “scene files” that can be visualized using Connectome Workbench software (see Supplementary Appendix S1). The neuroscience community at large can now exploit these resources for a variety of purposes, such as developing normative models, mapping new brain areas, analyzing connectomics, characterizing individual differences, and comparing with other suitably aligned datasets (either published or ongoing).

4Here we present a summary of the data quality and example results from the HCP 7T Retinotopy Dataset. The stimuli and analyses are detailed in the Methods and are described here very briefly. Each of 181 subjects participated in six 5-min pRF mapping runs. The stimuli comprised colorful object textures windowed through slowly moving apertures (Figure 1A). The colorful object textures were used because they produce high signal-to-noise ratio in higher level visual areas. The apertures were clockwise or counterclockwise rotating wedges, expanding or contracting rings, or bars that swept across the visual field in several directions (Figure 1B).

5The resource we provide with this paper is a large set of pRF model solutions. We define the pRF as the region of the visual field within which a visual stimulus elicits an increase in response from the pooled neural activity reflected in fMRI measurements, and can be summarized by the pRF's angle, eccentricity, and size (Figure 2A). The total dataset consists of 181 individual subjects and three group averages. The three group averages reflect two split halves of the subjects as well as all 181 subjects. For each of the 181 individuals and the three group averages, we solved three sets of models: one from the concatenation of all six runs (300 s per run, 1,800 time points), one from the first half of each run (150 s per run, 900 time points), and one from the second half of each run (150 s per run, 900 time points). For each subject or group average and for each of the three types of model fits, we obtained model solutions for the 91,282 cortical vertices and subcortical voxels (grayordinates spaced on average 2 mm apart). Each model solution yielded six numbers: angle, eccentricity, pRF size, and gain describing the pRF model, variance explained by the model, and mean signal intensity. Therefore in total, the pRF model solutions that we provide consist of 184 subjects (181 individuals and 3 group averages) × 91,282 grayordinates × 3 model fits × 6 quantities (Figure 2B). Individual subjects are referred to as S1–S181, the two split-half group averages are referred to as S182 and S183, and the full group average is referred to as S184.

6The particular form of the pRF model we employed assumes that each grayordinate's pRF is a two-dimensional (2-D) isotropic Gaussian and that contrast within the pRF is summed sublinearly according to a static power-law nonlinearity with exponent 0.05 (Kay, Winawer, et al. ). This can be expressed formally as where r(t) is the predicted stimulus-related time series, g is a gain parameter, S(t) is the stimulus aperture at time t, G is the 2-D isotropic Gaussian, n is an exponent parameter (n = 0.05), and h(t) is a canonical HRF. The subadditive exponent was used to obtain more accurate pRF solutions, but since it was fixed for all models we do not analyze it further. Note that the pRF sizes that we report are not the standard deviation of the 2-D Gaussian G, but rather the standard deviation of the Gaussian that describes the predicted response taking into account the effect of the subadditive exponent. The pRF size can be interpreted as the standard deviation of a 2-D Gaussian that characterizes how a grayordinate responds to point stimuli in the visual field (see Methods for details).

7The pRF model solutions can be used to make different types of inferences regarding visual response properties. First, if the pRF model successfully explains variance in the time series data for a grayordinate, this indicates that the grayordinate is visually responsive, but does not by itself imply spatial selectivity. For example, if a grayordinate responds with equal strength to a stimulus presented anywhere in the visual field, variance in its time series data can be explained by a pRF model that has very large spatial extent. Second, to be considered spatially selective, a grayordinate must not only be visually responsive but also exhibit larger responses to stimuli in some locations compared to others. To help assess spatial selectivity in the HCP dataset, we fit each grayordinate time series with a simple ON/OFF model that is sensitive only to the presence or absence of the stimulus, and compare the variance explained by this model to that explained by the full pRF model (see Supplementary Figure S1). Visual responsivity and spatial selectivity are properties of a single grayordinate. A third type of inference is retinotopic organization. This is a stronger claim that describes spatial selectivity at a larger scale: Retinotopic organization implies not only that single brain locations are spatially selective, but also that adjacent brain locations respond to nearby locations in the visual field, thereby producing smooth progressions of polar angle and/or eccentricity. In principle, a brain region might be spatially selective but not retinotopic if the spatial tuning of nearby brain locations is haphazard. In practice, this seems to be uncommon. Finally, a retinotopic map, or visual area, is generally considered to be a region of the brain that contains a representation of all or most of the contralateral visual hemifield in each hemisphere. In this paper, we make observations regarding retinotopic organization in the HCP dataset but do not attempt to resolve various ongoing controversies regarding human retinotopic maps (see Discussion).

8We first summarize pRF model solutions from group average S184 (which reflects all 181 individual subjects). Group-average results were obtained by taking the time series data from individual subjects (aligned using MSMAll to HCP's average cortical surface space fs_LR; see Methods), computing the across-subjects average of the time series data observed at each grayordinate, and then fitting a pRF model to the time series data at each grayordinate. For visualization, we map the results from fs_LR space to fsaverage space and plot the results on the fsaverage surface that has been inflated, spherized, and orthographically projected to a plane (Figure 3). We also provide visualizations of the results on the inflated fsaverage surface using dynamic rotating movies (Supplementary Movies S1 through S12).

9The effect of averaging the time series data across subjects differs across the cortex, depending on how well pRF parameters match between subjects given the MSMAll alignment. Prior work has shown that the V1–V3 maps have highly regular topography and are well aligned to measures of anatomy, such as surface curvature (Benson et al.,; Benson et al.,; Hinds et al., ) and myelination (Abdollahi et al., ), and to measures of function such as resting-state connectivity (Bock et al.,; Raemaekers et al., ). Therefore, these maps are likely to be well aligned across subjects, and averaging will preserve many of the features found in the maps of individual subjects. In particular, the angle and eccentricity maps show clear and expected patterns in V1–V3 (Figure 4, second and third columns), and the variance explained is greater than 75% (cyan regions in the fifth column of Figure 4). As expected, from the lower to upper bank of the calcarine sulcus, there is a smooth progression from the upper vertical meridian through the contralateral horizontal meridian to the lower vertical meridian (blue-cyan-green-yellow-red sweep in the angle color maps). The angle map reverses at the lips of the calcarine sulcus, with mirror-reversed and approximately quarter-field representations in the bordering dorsal and ventral V2 maps and dorsal and ventral V3 maps. As expected, the eccentricity map is in register across V1–V3, progressing from foveal to peripheral representations from near the occipital pole toward medial and anterior directions (blue-magenta-red-yellow-green progression in the eccentricity colormap). The pRF size map has some of the same features of the eccentricity map, exhibiting smaller sizes near the occipital pole and larger sizes in the midperipheral regions of V1–V3. However, in the more peripheral portions of the maps, the size estimates are smaller than predicted from eccentricity due to stimulus edge effects (blue rim around the anterior/medial edge of the V1–V3 maps; see Discussion).

10In cortical locations where pRF parameters are variable across subjects (even after registration using MSMAll), the group-average results will preserve less of the detail from individual subjects. Nonetheless, there is a large amount of structure in the group-average results beyond V1–V3, and some clear patterns are evident. The angle maps show the expected progression from upper to lower field ventral to V3, and from lower to upper field dorsal to V3 (Locations 1 and 2 in Figure 4, top row), consistent with measurements of ventral (Kastner et al.,; McKeefry, Watson, Frackowiak, Fong, & Zeki,; Wade, Brewer, Rieger, & Wandell, ) and dorsal (Press, Brewer, Dougherty, Wade, & Wandell,; Tootell et al., ) occipital cortex. The eccentricity map also shows clear large-scale organization throughout large expanses of parietal and temporal cortex. One feature of the eccentricity maps is multiple distinct, foveal representations: In addition to the foveal representation in V1–V3 at the occipital pole, the eccentricity maps show distinct foveal representations in ventral temporal cortex and parietal cortex (Locations 3 and 4 in Figure 4), consistent with many prior studies (Swisher, Halko, Merabet, McMains, & Somers,; Tootell et al.,; Wade et al.,; Wandell, Brewer, & Dougherty, ). Near both of these distinct foveal representations, there are foveal to peripheral progressions along the lateral to medial direction.

11The pRF size map also shows a variety of large-scale patterns. In ventral temporal cortex, there is a small-to-large size gradient from the fusiform gyrus to the collateral sulcus (Locations 5 and 6 in Figure 4). These regions roughly correspond to the locations of face-selective and place-selective cortex (Epstein & Kanwisher,; Grill-Spector & Weiner,; Kanwisher, McDermott, & Chun, ). More generally, pRF sizes tend to be larger outside V1–V3, as expected from both single-unit and fMRI measurements (Dumoulin & Wandell,; Maunsell & Van Essen,; Smith, Singh, Williams, & Greenlee,; Tootell et al., ). Finally, the variance explained map shows that robust signals occur not only within V1–V3 but also in higher level areas. Variance explained is above 50% in several regions ventral, lateral, and dorsal to the V1–V3 maps, including much of ventral temporal cortex and the intraparietal sulcus (Locations 7 and 8 in Figure 4). Furthermore, for nearly all the cortical locations that survive the variance explained threshold, pRF model parameters are highly reliable. This is confirmed by two types of split-half analysis: First, the data averaged across all 181 subjects (S184) were split into two halves by time, with one dataset comprising time series from the first half of each of the six runs, and a second dataset comprising time series from the second half of each of the six runs. Second, the data were split into two halves by subject, with one dataset reflecting averaging time series data across 91 subjects (S182), and a second dataset reflecting averaging time series data across the remaining 90 subjects (S183). Both split-half analyses indicate high reliability of pRF parameters (results not shown; pRF model solutions available online).

12Many features of the group-average results are in good agreement with recently published parcellations of visual areas, particularly near the posterior occipital pole (Figure 5). We compare the group-average pRF results to two atlases made using different methods: the Wang et al. maximum probability atlas and the Glasser et al. (Glasser, Coalson, et al., ) multimodal parcellation of cortex. The Wang et al. maximum probability atlas includes 25 regions of interest (ROIs) per hemisphere whose boundaries are derived from the anatomically aligned overlap of manually labeled visual areas in individual subjects. Ten of these in posterior cortex are clearly aligned with expected features of the angle maps: V1v/V1d, V2v/V2d, V3v/V3d, V3A, V3B, LO-1, and hV4. In each of these 10 ROIs, one or more borders lie on an angle reversal. For example, the V1d/V2d border lies on a lower-field angle reversal, and the V1v/V2v border lies on an upper-field angle reversal. The agreement between the retinotopic features of the HCP dataset and the borders from the Wang et al. atlas is remarkably good despite the atlas reflecting different subjects and different experimental and analysis methods. Other maps such as LO-2, TO-1/2, IPS maps, VO-1/2, and PHC-1/2 show contralateral representations but not clear progressions of angle, in part due to blurring from group averaging. For example, imperfect alignment of the cortex might result in pRF parameters from one subject's upper visual field being averaged with those of another subject's lower visual field. The Glasser et al. parcellation was generated using a semi-automated, supervised approach applied to multimodal neuroimaging data (representing architecture, connectivity, function, and visuotopic organization). For the Glasser et al. parcellation, several areas are well aligned with features of the retinotopic maps, particularly V1, V2, V3, V4, and V3A. In several map clusters in the Wang et al. atlas, there are clear eccentricity gradients: The IPS0–2 maps show a clear foveal-to-peripheral gradient along the medial-to-lateral direction, as do the V1–V3 maps and the VO-1/2 maps. In the Glasser et al. atlas, several areas fall within iso-eccentricity regions. For example, the areas PH and TE2p are clearly foveal, whereas the adjacent fusiform face complex is more peripheral, with a sharp change in eccentricity along the border between these areas. Consistent with our definitions in the previous sections, the eccentricity gradients indicate clear retinotopic organization in these anterior areas; however, we do not try to draw conclusions in this paper regarding whether there are complete maps of the visual field or the most appropriate way to parcellate cortex into distinct areas.

13The HCP 7T Retinotopy Dataset includes subcortical results in addition to cortical results. Several subcortical nuclei have retinotopic maps that have been previously measured using fMRI (Arcaro, Pinsk, & Kastner,; Cotton & Smith,; DeSimone, Viviano, & Schneider,; Katyal, Zughni, Greene, & Ress,; Schneider & Kastner,; Schneider, Richter, & Kastner, ). The subcortical fMRI data were aligned using FNIRT (FSL Nonlinear Image Registration Tool) nonlinear volume registration based on T1-weighted image intensities (Glasser et al., ). In contrast to cortex, subcortical structures are not easily represented as 2-D surfaces, and hence it is more difficult to visualize complete maps. Nonetheless, slices through subcortical structures reveal clear, high-quality pRF model solutions in the group-average dataset (Figure 6). In particular, we see expected structure in visual nuclei such as the lateral geniculate nucleus (LGN), superior colliculus (SC), and ventral pulvinar (vPul1/2). Within these regions, there are clear representations of the contralateral visual field. As expected, the visual field maps of the LGN and pulvinar are both inverted with smooth progressions from the upper visual field located ventrally to the lower visual field located dorsally. In the superior colliculus, there is a smooth progression from the upper visual field (anterior and medial) to the lower visual field (posterior and lateral).

14In addition to group-average results, we also computed pRF model solutions for the 181 individual subjects. We summarize results in several ways, including quantifying the amount of variance explained by the pRF model, inspecting maps in individual subjects, and assessing within-subject reliability of pRF parameters. These analyses reveal that overall data quality is high.

15We quantified variance explained by the pRF model within atlas-defined ROIs. We defined one ROI as the union of the 50 maps found in the Wang et al. maximum probability atlas (25 maps per hemisphere) and a second ROI as the union of the V1–V3 maps from the same atlas. The V1–V3 ROI is a subset of the larger ROI. Because these ROIs are defined based on group-average anatomy, they do not necessarily conform to each individual subject's retinotopic maps, but they provide a simple objective method for region definition. Within the union of the 50 maps, we computed for each subject the median variance explained across grayordinates, yielding one number per subject. The median of this number across the 181 subjects was 17% (Figure 7A). Within just the V1–V3 maps, the median of the median variance explained was substantially higher, at 44%. For comparison, we estimate that for grayordinates not sensitive to the experimental paradigm, the variance explained by the pRF model is less than 1%. This can be seen by inspecting the large peak in the histogram of variance explained across all grayordinates from all individual subjects (Figure 7C).

16For map visualization, we selected three representative subjects: the subjects at the 25th, 50th, and 75th percentiles with respect to median variance explained across regions in the Wang et al. atlas (see red lines in Figure 7A). For simplicity we show only the left hemisphere, and we replot the group-average results for comparison. The three depicted subjects have clear retinotopic maps in occipital cortex, as seen in the angle and eccentricity results (Figure 7C). In each subject, the angle maps reveal the boundaries of V1–V3, and the eccentricity maps are in register across visual field maps around the occipital pole. The locations of the V1–V3 boundaries differ slightly across the subjects, as seen by comparing the angle reversals and the V1–V3 boundary lines that were drawn based on the group-average results. This suggests that even after alignment using state-of-the-art algorithms guided by folding and areal features (MSMAll), there is residual misalignment of retinotopic maps in some subjects. A few subjects showed strikingly atypical retinotopy, such as a “forked” representation of the lower vertical meridian representation running across V2d and V3d of the group average in subjects S80 (HCP ID 198653) and S138 (HCP ID 644246) (see Figure S2 in Van Essen & Glasser, and https://balsa.wustl.edu/ZLV7).

17Beyond V1–V3, several of the features we noted in the group-average results are also generally evident in the individual subjects. For example, the angle maps show a lower-field representation ventral to V3 and an upper-field representation dorsal to V3 (Locations 1 and 2 in Figure 7). There are also distinct foveal representations in parietal and temporal cortex (Locations 3 and 4), and pRF size gradients in ventral cortex (Locations 5 and 6). Because variance explained is generally lower for individual subjects compared to the group average, there are some regions in which the group average may provide useful information that is absent in individual subjects (e.g., Location 8). By visual inspection, the overall map quality appears comparable across the three subjects. Since these subjects span the central 50% of variance explained (as detailed previously), this suggests that most of the subjects in the HCP 7T Retinotopy Dataset have good data quality. Additional aspects of individual variability can be readily inspected by scrolling through polar angle and eccentricity maps for all 181 individual subjects in the downloadable Connectome Workbench “scene” files (see Supplementary Appendix S1).

18To quantify reliability of pRF parameters for individual subjects, we compared parameter estimates across split-halves of the data. We binned cortical grayordinates into four large ROIs that comprise distinct subsets of the regions in the Wang et al. atlas: posterior (V1–V3), dorsal (V3A/B, IPS0–5), lateral (LO-1/2, TO-1/2), and ventral (VO-1/2, PHC-1/2). We then aggregated grayordinates within each of these ROIs across subjects, and computed 2-D histograms comparing parameter estimates across the two model fits (first half of each run; second half of each run).

19Angle estimates were highly reliable across splits for all four ROIs, indicated by the high density along the diagonals (Figure 8, top row). In addition to demonstrating within-subject reliability, these histograms highlight the fact that angles near the vertical meridian (90° and 270°) are less represented than other angles, an effect observed in many prior studies (Arcaro, McMains, Singer, & Kastner,; Kastner et al.,; Larsson & Heeger,; Mackey, Winawer, & Curtis,; Silver, Ress, & Heeger,; Swisher et al., ). This effect is likely due, in part, to the pooling intrinsic to the fMRI signal. Assuming that neural receptive field centers are all contralateral, the upper and lower vertical meridians are boundaries of the distribution of centers, and the average polar angle of neurons within a voxel is not expected to reach the vertical meridians. The eccentricity histograms (Figure 8, middle row) also show a high degree of reliability, with density highest on and near the diagonal in all four ROIs. Note that the dorsal ROI, while reliable, is more foveally biased than other ROIs. Nonetheless, as indicated in both the maps (Figures 4 and 7) and the reliability plots (Figure 8), the dorsal regions contain eccentricities spanning 0° to 8°. Finally, the size estimates were also fairly reliable, though less so than the angle and eccentricity estimates. In agreement with the maps (Figures 4 and 7), posterior maps generally contain the smallest pRFs, with few pRF sizes larger than 3°. The high reliability of pRF parameters across data splits supports the interpretation that grayordinates in not only V1–V3 but also dorsal, lateral, and ventral higher extrastriate regions exhibit spatial selectivity and not mere visual responsivity. If grayordinates responded indiscriminately to a visual stimulus presented anywhere in the visual field, pRF parameters would not exhibit such reliable tuning for specific pRF parameter values.

20In this study, we have described the HCP 7T Retinotopy Dataset and the results of fitting pRF models to all 181 individual subjects as well as the group average. To facilitate quantification of model reliability, all datasets were analyzed using split halves in addition to the full dataset. In addition to the pRF model solutions, we also make available the stimuli and analysis code used to solve the models. This allows the research community to reproduce our analyses and/or reanalyze the time series data using different techniques. The analyses we conducted are computationally intensive, involving three independent nonlinear optimizations for each grayordinate time series. This resulted in approximately 50 million model fits that necessitated the use of a large-scale compute cluster. By providing the parameters of the solved models, we substantially lower the barrier to entry for scientists to make use of the dataset.

21Although researchers frequently collect, and occasionally make public, retinotopy datasets, such datasets have generally included no more than 20 subjects (e.g., Benson et al., ). To the best of our knowledge, the HCP 7T Retinotopy Dataset is the largest publicly available dataset by an order of magnitude. In addition to containing many subjects, retinotopic maps are derived from six fMRI runs (a total of 30 min of data), making this dataset large both in terms of number of subjects as well as amount of data per subject. Finally, the data were acquired at ultrahigh magnetic field strength (7T), providing enhanced signal-to-noise ratio and high spatial and temporal resolution (1.6-mm isotropic voxels, 1-s temporal sampling using multiband data acquisition). The advantages of the dataset are clear. In individual subjects, there are reliable results, even beyond striate and extrastriate cortex. At the group level, the massive averaging of subjects reveals signals in regions of cortex (such as the inferior frontal sulcus) where conventional datasets typically have low signal that can be difficult to distinguish from noise.

22Though the dataset has clear value, it is also important to understand its limitations and take these into account when interpreting the data. There are several technical issues; we mention a few here, but refer to the Methods for a fuller description. The stimulus size extended to an eccentricity of 8° of visual angle, and so representations of the far periphery are not well measured. Because of cortical magnification of the central visual field, robust signals are found in about half of the surface area of V1 (see Figure 5). Edge effects arise for grayordinates whose pRF centers are near the limit of the stimulus extent: These grayordinates are likely to have underestimates of pRF size and a displaced pRF center. Model solutions are somewhat discretized, reflecting the influence of the first-stage grid fit. Model solutions were constrained to have a nonnegative gain factor; this may not be appropriate for studying brain regions that exhibit BOLD signal decreases in response to visual stimulation. Finally, there is an inherent limitation related to the fact that we analyzed the data using one specific pRF model with a particular code implementation. The motivation of this paper is to use established tools and models to generate a high-quality retinotopy resource, but some scientific questions will require additional modeling work. Exploring other pRF models (such as a difference-of-Gaussians pRF model (Zuiderbaan, Harvey, & Dumoulin, ), an anisotropic Gaussian pRF model (Merkel, Hopf, & Schoenfeld,; Silson, Reynolds, Kravitz, & Baker, ), or pRF models with flexible shapes (Greene, Dumoulin, Harvey, & Ress,; Lee, Papanikolaou, Logothetis, Smirnakis, & Keliris, ) and carefully evaluating model accuracy (Kay, Winawer, et al., ) may be important to answer such questions.

23In addition to modeling choices, it is important to also consider stimulus selectivity and task effects. In particular, there may be parts of the brain that show retinotopic organization given an appropriate stimulus and task but not in all retinotopic experimental paradigms. In the experiment used for the HCP dataset, the images within the moving apertures changed 15 times per second. Cortical areas with low-pass temporal tuning might not be effectively driven by these stimuli (Liu & Wandell, ). Other areas, for example in the dorsal attention network, might respond strongly only when stimuli are attended (Mackey et al., ); in the HCP experiment, subjects attended to the fixation location. Hence, a lack of spatial selectivity or retinotopic organization in this dataset, even with the large number of subjects, should not be taken as definitive evidence that a brain area is not spatially selective or retinotopically organized.

24The neuroscientific interpretation of the pRF results must also be done carefully. Whereas in visual cortex, there is clear interpretation of pRF models in terms of visually responsive population receptive fields, in other parts of the brain, it may be possible to obtain good pRF fits but for different reasons. For example, it is possible that a cortical region indexing cognitive difficulty exhibits response increases when the stimulus is near the fovea because at these points in the experiment, the stimulus is more likely to interfere with the fixation task performed by the subjects. In such a case, the existence of a pRF model solution does not imply visually driven activity in the conventional sense.

25The group-average datasets (S182–S184) provide a useful summary of the overall organization of visually responsive cortex. Because of the large number of subjects that are averaged together and the improved intersubject alignment methods, these datasets contain very high signal-to-noise level. However, there are several caveats to interpreting these data. Most significantly, unlike in individual subjects, the quality of pRF model fits is influenced by the quality of the alignment used to generate these averages. In particular, the quality of the fits in V1, V2, and V3 of the group average appears to be very high; this is due both to the fact that visually evoked signals in V1–V3 are particularly robust, as well as the fact that these cortical areas are less variable in their locations and internal topographic structure than higher visual areas (Benson et al., ). As one moves from striate to extrastriate cortex and beyond, coverage of the visual field becomes less and less complete. The effect of averaging can be appreciated by comparing the group-average data to the individual data (see Figure 7). In the group-average data, the V1 maps sweep out nearly 180° of polar angle from the lower vertical meridian (deep red) to the upper vertical meridian (deep blue); the V3 maps do not quite reach the vertical meridians, and V3A (Location 2 in the angle map) reaches only about 45° (cyan as opposed to blue). In contrast, the individual-subject maps, though noisier, often exhibit more representation of the vertical meridian in V3 and V3A (deep blue and red in the angle maps of Figure 7). Indeed, previous work has found that for most visual areas beyond V1–V3, the overlap among subjects decreases substantially (Wang et al., ). One analysis that could help shed light on these issues is to examine intersubject variability in pRF parameters as a function of cortical location.

26This rich dataset has a wide range of uses. It provides the basis for further analysis of other HCP data; for example, the pRF solutions for an individual subject can be used to determine visual ROIs that could then be used to analyze or validate other HCP measures. Some example applications include the following: (a) The retinotopy dataset can be used for comparison with the HCP's multimodal parcellation (Glasser, Coalson, et al.,; see Figure 5). We have shown that the group-average results approximately agree with portions of the parcellation, but we did not compare individual-subject results to the group-average parcellation or to the individual-subject parcellations that were generated using an areal classifier algorithm (Glasser, Coalson, et al., ). (b) Identifying visual ROIs and pRF properties within the ROIs can be used in conjunction with resting-state data (Van Essen et al., ) to test hypotheses about how maps relate to functional connectivity. (c) The pRF model solutions can be used in conjunction with the working memory dataset (Barch et al., ) to study the role of visual cortex in working memory. Many more such applications (e.g., combining retinotopy with the 7T movie data) are possible.

27The visuotopic mapping in the Glasser, Coalson, et al. parcellation was based on resting-state fMRI correlations measured across the entire visual field representation. This enabled mapping the full extent of visuotopic areas, but does not provide explicit measurement of specific eccentricities or polar angles within each map. Hence, the current 7T retinotopic maps and the visuotopic organization derivable from resting-state data represent complementary and potentially synergistic information.

28The HCP 7T Retinotopy Dataset also has a great deal of standalone value, owing to the very large number of subjects. Any examination of the relationship between anatomy and function benefits from having many subjects to characterize the extent of intersubject structure–function variability in an anatomically normalized format. Averaging retinotopic time series data across a large number of subjects has revealed that large swaths of cortex not typically studied by vision scientists show evidence of retinotopic organization (see Figure 4); many of these regions would not have clear signals in smaller sample sizes.

29Despite 25 years of measuring retinotopic maps with fMRI, a number of disagreements concerning map organization remain unresolved. For example, there are two different proposals for the organization of the V4 map and neighboring regions. The hV4/VO proposal consists of a single hV4 hemifield map on the ventral surface, with several additional hemifield maps located more anterior (Arcaro et al.,; McKeefry & Zeki,; Wade et al.,; Winawer & Witthoft, ). The V4/V8 proposal involves a different arrangement, with the V4 map split into dorsal and ventral arms (Hansen, Kay, & Gallant,; Sereno et al., ) and with a V8 hemifield map adjacent to ventral V4 (Hadjikhani, Liu, Dale, Cavanagh, & Tootell,; Hadjikhani & Tootell,; Tootell & Hadjikhani, ). These two different proposals are implicit in the parcellations shown in Figure 5 from Wang et al. and Glasser, Coalson, et al.. There are a number of additional unresolved questions regarding map organization, including the number and arrangement of maps in the vicinity of MT (Amano, Wandell, & Dumoulin,; Kolster, Peeters, & Orban, ). In fact, beyond the V1–V3 maps, there is likely no retinotopic map that is universally agreed upon by researchers in the field. Part of this is due to the challenge of interpreting complex spatial data; other disagreements might stem from differences across datasets (e.g., due to differences in MRI acquisition methods, stimuli, analysis approaches, and subjects). Indeed, even within V1–V3, there are unresolved questions about retinotopic organization, such as how precisely the maps align with anatomical landmarks and whether some individuals have maps that qualitatively differ from the typical pattern (see supplemental information in Van Essen & Glasser,; https://balsa.wustl.edu/ZLV7). We believe that the HCP 7T Retinotopy Dataset provides a unique opportunity to adjudicate among competing hypotheses about the organization of retinotopic representations in the human brain. Future work, perhaps exploiting automated, objective atlas-based fitting procedures (e.g., Benson & Winawer, ), could help evaluate how well different proposals are supported by the data.

30The visual system is one of the primary functional systems of the human brain, and the resources provided in this paper represent an important step towards more fully characterizing its fundamental organization. The authors believe that the present measurements fill a critical role, both for answering novel scientific questions and for establishing baselines and hypotheses for new experiments. To this end, we have put effort into making all data and analyses fully public and well-documented, and we hope that other researchers will find this dataset enlightening and useful.