Author response:
Reviewer #1 (Public review):
Summary:
The authors examine the neural correlates of face recognition deficits in individuals with Developmental Prosopagnosia (DP; 'face blindness'). Contrary to theories that poor face recognition is driven by reduced spatial integration (via smaller receptive fields), here the authors find that the properties of receptive fields in face-selective brain regions are the same in typical individuals vs. those with DP. The main analysis technique is population Receptive Field (pRF) mapping, with a wide range of measures considered. The authors report that there are no differences in goodness-of-fit (R2), the properties of the pRFs (neither size, location, nor the gain and exponent of the Compressive Spatial Summation model), nor their coverage of the visual field. The relationship of these properties to the visual field (notably the increase in pRF size with eccentricity) is also similar between the groups. Eye movements do not differ between the groups.
Strengths:
Although this is a null result, the large number of null results gives confidence that there are unlikely to be differences between the two groups. Together, this makes a compelling case that DP is not driven by differences in the spatial selectivity of face-selective brain regions, an important finding that directly informs theories of face recognition. The paper is well written and enjoyable to read, the studies have clearly been carefully conducted with clear justification for design decisions, and the analyses are thorough.
Weaknesses:
One potential issue relates to the localisation of face-selective regions in the two groups. As in most studies of the neural basis of face recognition, localisers are used to find the face-selective Regions of Interest (ROIs) - OFA, mFus, and pFus, with comparison to the scene-selective PPA. To do so, faces are contrasted against other objects to find these regions (or scenes vs. others for the PPA). The one consistent difference that does emerge between groups in the paper is in the selectivity of these regions, which are less selective for faces in DP than in typical individuals (e.g., Figure 1B), as one might expect. 6/20 prosopagnosic individuals are also missing mFus, relative to only 2/20 typical individuals. This, to me, raises the question of whether the two groups are being compared fairly. If the localised regions were smaller and/or displaced in the DPs, this might select only a subset of the neural populations typically involved in face recognition. Perhaps the difference between groups lies outside this region. In other words, it could be that the differences in prosopagnosic face recognition lie in the neurons that are not able to be localised by this approach. The authors consider in the discussion whether their DPs may not have been 'true DPs', which is convincing (p. 12). The question here is whether the regions selected are truly the 'prosopagnosic brain areas' or whether there is a kind of survivor bias (i.e., the regions selected are normal, but perhaps the difference lies in the nature/extent of the regions. At present, the only consideration given to explain the differences in prosopagnosia is that there may be 'qualitative' differences between the two (which may be true), but I would give more thought to this.
We acknowledge that face-selective ROIs in DPs, relative to controls, may be smaller, less selective, or altogether missing when traditional methods of localization with fixed thresholds are used (Furl et al, 2011). For this reason - to circumvent potential survivor bias and ensure ROI voxel counts across participants are equated - we used a method of ROI definition whereby each subject’s individual statistical map from the localizer was intersected with a generously-sized group mask for each ROI and the top 20% most category-selective voxels were retained for the pRF analysis (Norman-Haignere et al., 2013; Jiahui et al., 2018). This means that the raw number of voxels per ROI was equal across all participants with respect to the common group space, thereby ensuring a fair comparison even in cases where one group shows diminished category-selectivity. The details of the ROI definition are provided in the Methods at the end of the manuscript. To ensure readers understand our approach, we will also make more explicit mention of this in the main body of the manuscript.
With regard to the question of whether face-selective ROIs may be displaced in DPs compared to controls, previous work from the senior author’s lab (Jiahui et al., 2018) shows that, despite exhibiting weaker activations, the peak coordinates of significant clusters in DPs occupy very similar locations to those of controls. And, even if there were indeed slight displacements of face-selective ROIs for some subjects, the group-defined masks used in the present analysis were large enough to capture the majority of the top voxels. In the supplemental materials section, we will include a diagram of the group masks used in our study.
The reviewer here also points out that more DPs than controls were missing the mFUS region (6/20 DPs vs 2/20 controls; Figure 1C). However, ‘missing’ in this context was not based on face-selectivity but rather a lack of retinotopic tuning. PRFs were fit to all voxels within each ROI - with all subjects starting out with equal voxel counts - and thereafter, voxels for which the variance explained by the pRF model was below 20% were excluded from subsequent analysis. We decided that any ROI with fewer than 10 voxels remaining after thresholding on the pRF fit should be deemed ‘missing’ since we considered the amount of data insufficient to reliably characterize the region’s retinotopic profile. While it may be somewhat interesting that four more DPs than controls were ‘missing’ left mFUS, using this particular set of decision criteria, it is important to keep in mind that left mFUS was just one of six face-selective regions under study. The other five regions, many of which evinced strong fits by the pRF model, were represented comparably in DPs and controls and showed high similarity in the pRF parameters. Furthermore, across most participants, mFUS exhibited a low proportion of retinotopically modulated voxels (defined as voxels with pRF R squared greater than 20%, see Figure 1D). A follow-up analysis showed that the count of voxels surviving pRF R squared thresholding in left mFUS was not significantly correlated with mean pRF size (r(30)=0.23, t=1.28, p=0.21) indicating that the greater exclusion of DPs in this region is unlikely to have biased the group’s average pRF size.
The discussion considers the differences between the current study and an unpublished preprint (Witthoft et al, 2016), where DPs were found to have smaller pRFs than typical individuals. The discussion presents the argument that the current results are likely more robust, given the use of images within the pRF mapping stimuli here (faces, objects, etc) as opposed to checkerboards in the prior work, and the use of the CSS model here as opposed to a linear Gaussian model previously. This is convincing, but fails to address why there is a lack of difference in the control vs. DP group here. If anything, I would have imagined that the use of faces in mapping stimuli would have promoted differences between the groups (given the apparent difference in selectivity in DPs vs. controls seen here), which adds to the reliability of the present result. Greater consideration of why this should have led to a lack of difference would be ideal. The latter point about pRF models (Gaussian vs. CSS) does seem pertinent, for instance - could the 'qualitative' difference lead to changes in the shape of these pRFs in prosopagnosia that are better characterised by the CSS model, perhaps? Perhaps more straightforwardly, and related to the above, could differences in the localisation of face-selective regions have driven the difference in prior work compared to here?
We agree that the use of high-level mapping stimuli (including faces) adds to the reliability of the present results for DPs and could have further emphasized differences between the groups if true differences did, in fact, exist. We speculate on the extent to which the type of mapping stimuli and various other methodological factors (e.g. stimulus size, aperture design, pRF model) could have explained the divergent findings in our study versus that of Witthoft et al. (2016) in the section of the Discussion titled, “What factors may have contributed to the different results for the present study and Witthoft et al. (2016)”. In brief, our use of more colorful, naturalistic stimuli targeting higher-level visual areas elicited better model fits than the black and white checkerboard pattern used by Witthoft et al. (2016). The CSS model we used is better suited for higher-level regions and makes fewer assumptions than the linear pRF model. The field of view of our stimulus was smaller but still relevant for real-world perception of faces. Finally, our aperture design and longer run length likely also improved reliability. Overall, these methodological improvements, along with our larger sample size, provide stronger evidence for our findings. These are our best attempts to make sense of the divergent findings, but it is not possible to come to a definitive explanation. Examples abound of exaggerated or spurious effects from small-scale studies that ultimately fail to replicate in the related field of dyslexia research (Jednorog et al., 2015; Ramus et al., 2018) and neuroimaging research more generally (Turner et al., 2018; Poldrack et al., 2017). Sometimes there are clear explanations for a lack of replicability (e.g. software bugs, overly flexible preprocessing methods, etc.), but many times the real reason cannot be determined.
Regarding the type of pRF model deployed, our use of a non-linear exponent (versus a linear model as in the Witthoft et al. (2016) preprint) is unlikely to explain the similarity we observed between the groups in terms of pRF size. Specifically, the groups did not show substantial differences in the exponent by ROI, as seen in Figure 1E, so the use of a linear model should, in theory, produce similar outcomes for the two groups. We will mention this point in the main text.
Finally, the lack of variations in the spatial properties of these brain regions is interesting in light of the theories that spatial integration is a key aspect of effective face recognition. In this context, it is interesting to note the marked drop in R2 values in face-selective regions like mFus relative to earlier cortex. The authors note in some sense that this is related to the larger receptive field size, but is there a broader point here that perhaps the receptive field model (even with Compressive Spatial Summation) is simply a poor fit for the function of these areas? Could it be that these areas are simply not spatial at all? A broader link between the null results presented here and their implications for theories of face recognition would be ideal.
The weaker pRF fits found in mFUS, to us, raise the question of whether there is a more effective pRF stimulus for these more anterior regions. For example, it might be possible to obtain higher and more reliable responses there using single isolated faces (Cf. Kay, Weiner, Grill-Spector, 2015). More broadly, though, we agree that it is important to acknowledge that the receptive field model might ultimately be a coarse and incomplete characterization of neural function in these areas. As the other reviewer suggests, one possibility is that other brain processes (e.g. functional or structural connectivity between ROIs) may give rise to holistic face processing in ways that are not captured by pRF properties.
Reviewer #2 (Public review):
Summary:
This is a well-conducted and clearly written manuscript addressing the link between population receptive fields (pRFs) and visual behavior. The authors test whether developmental prosopagnosia (DP) involves atypical pRFs in face-selective regions, a hypothesis suggested by prior work with a small DP sample. Using a larger cohort of DPs and controls, robust pRF mapping with appropriate stimuli and CSS modeling, and careful in-scanner eye tracking, the authors report no group differences in pRF properties across the visual processing hierarchy. These results suggest that reduced spatial integration is unlikely to account for holistic face processing deficits in DP.
Strengths:
The dataset quality, sample size, and methodological rigor are notable strengths.
Weaknesses:
The primary concern is the interpretation of the results.
(1) Relationship between pRFs and spatial integration
While atypical pRF properties could contribute to deficits in spatial integration, impairments in holistic processing in DPs are not necessarily caused by pRF abnormalities. The discussion could be strengthened by considering alternative explanations for reduced spatial integration, such as altered structural or functional connectivity in the face network, which has been reported to underlie DP's difficulties in integrating facial features.
We agree the Discussion section could benefit from mentioning that alterations to other neural mechanisms, besides pRF organization, could produce deficits in holistic processing. This could take the form of altered functional connectivity (Rosenthal et al., 2017; Lohse et al., 2016; Avidan et al., 2014) or altered structural connectivity (Gomez et al., 2015; Song et al., 2015)
(2) Beyond the null hypothesis testing framework
The title claims "normal spatial integration," yet this conclusion is based on a failure to reject the null hypothesis, which does not justify accepting the alternative hypothesis. To substantiate a claim of "normal," the authors would need to provide analyses quantifying evidence for the absence of effects, e.g., using a Bayesian framework.
We acknowledge that, using frequentist statistical methods, failing to reject the null hypothesis is not sufficient to claim equivalence. For the revision, we will look into additional analyses that could quantify evidence for the null hypothesis. And we will adjust the wording of the title in this regard.
(3) Face-specific or broader visual processing
Prior work from the senior author's lab (Jiahui et al., 2018) reported pronounced reductions in scene selectivity and marginal reductions in body selectivity in DPs, suggesting that visual processing deficits in DPs may extend beyond faces. While the manuscript includes PPA as a high-level control region for scene perception, scene selectivity was not directly reported. The authors could also consider individual differences and potential data-quality confounds (tSNR difference between and within groups, several obvious outliers in the figures, etc). For instance, examining whether reduced tSNR in DPs contributed to lower face selectivity in the DP group in this dataset.
Thank you for this suggestion - we will compare tSNR between the groups as a measure of data quality and we will include these comparisons. A preliminary look indicates that both groups possessed similar distributions of tSNR across many of the face-selective regions investigated here.
(4) Linking pRF properties to behavior
The manuscript aims to examine the relationship between pRF properties and behavior, but currently reports only one aspect of pRF (size) in relation to a single behavioral measure (CFMT), without full statistical reporting:
"We found no significant association between participants' CFMT scores and mean pRF size in OFA, pFUS, or mFUS."
For comprehensive reporting, the authors could examine additional pRF properties (e.g., center, eccentricity, scaling between eccentricity and pRF size, shape of visual field coverage, etc), additional ROIs (early, intermediate, and category-selective areas), and relate them to multiple behavioral measures (e.g., HEVA, PI20, FFT). This would provide a full picture of how pRF characteristics relate to behavioral performance in DP.
We will report the full statistical values (r, p) for the (albeit non-significant) relationship between CFMT score and pRF size - thank you for bringing that to our attention. Additionally, we will add other analyses assessing the relationship between a wider array of pRF measures and the other behavioral tests administered to provide a more comprehensive picture of the relation between pRFs and behavior.
References:
Avidan, G., Tanzer, M., Hadj-Bouziane, F., Liu, N., Ungerleider, L. G., & Behrmann, M. (2014). Selective Dissociation Between Core and Extended Regions of the Face Processing Network in Congenital Prosopagnosia. Cerebral Cortex, 24(6), 1565–1578. https://doi.org/10.1093/cercor/bht007
Furl, N., Garrido, L., Dolan, R. J., Driver, J., & Duchaine, B. (2011). Fusiform gyrus face selectivity relates to individual differences in facial recognition ability. Journal of Cognitive Neuroscience, 23(7), 1723–1740. https://doi.org/10.1162/jocn.2010.21545
Gomez, J., Pestilli, F., Witthoft, N., Golarai, G., Liberman, A., Poltoratski, S., Yoon, J., & Grill-Spector, K. (2015). Functionally Defined White Matter Reveals Segregated Pathways in Human Ventral Temporal Cortex Associated with Category-Specific Processing. Neuron, 85(1), 216–227. https://doi.org/10.1016/j.neuron.2014.12.027
Jednoróg, K., Marchewka, A., Altarelli, I., Monzalvo Lopez, A. K., van Ermingen-Marbach, M., Grande, M., Grabowska, A., Heim, S., & Ramus, F. (2015). How reliable are gray matter disruptions in specific reading disability across multiple countries and languages? Insights from a large-scale voxel-based morphometry study. Human Brain Mapping, 36(5), 1741–1754. https://doi.org/10.1002/hbm.22734
Jiahui, G., Yang, H., & Duchaine, B. (2018). Developmental prosopagnosics have widespread selectivity reductions across category-selective visual cortex. Proceedings of the National Academy of Sciences of the United States of America, 115(28), E6418–E6427. https://doi.org/10.1073/pnas.1802246115
Kay, K. N., Weiner, K. S., Kay, K. N., & Weiner, K. S. (2015). Attention Reduces Spatial Uncertainty in Human Ventral Temporal Cortex Attention Reduces Spatial Uncertainty in Human Ventral Temporal Cortex. Current Biology, 25(5), 595–600. https://doi.org/10.1016/j.cub.2014.12.050
Lohse, M., Garrido, L., Driver, J., Dolan, R. J., Duchaine, B. C., & Furl, N. (2016). Effective connectivity from early visual cortex to posterior occipitotemporal face areas supports face selectivity and predicts developmental prosopagnosia. Journal of Neuroscience, 36(13), 3821–3828. https://doi.org/10.1523/JNEUROSCI.3621-15.2016
Norman-Haignere, S., Kanwisher, N., & McDermott, J. H. (2013). Cortical pitch regions in humans respond primarily to resolved harmonics and are located in specific tonotopic regions of anterior auditory cortex. Journal of Neuroscience, 33(50), 19451–19469. https://doi.org/10.1523/JNEUROSCI.2880-13.2013
Poldrack, R. A., Baker, C. I., Durnez, J., Gorgolewski, K. J., Matthews, P. M., Munafò, M. R., Nichols, T. E., Poline, J. B., Vul, E., & Yarkoni, T. (2017). Scanning the horizon: Towards transparent and reproducible neuroimaging research. Nature Reviews Neuroscience, 18(2), 115–126. https://doi.org/10.1038/nrn.2016.167
Ramus, F., Altarelli, I., Jednoróg, K., Zhao, J., & Scotto di Covella, L. (2018). Neuroanatomy of developmental dyslexia: Pitfalls and promise. Neuroscience and Biobehavioral Reviews, 84(July 2017), 434–452. https://doi.org/10.1016/j.neubiorev.2017.08.001
Rosenthal, G., Tanzer, M., Simony, E., Hasson, U., Behrmann, M., & Avidan, G. (2017). Altered topology of neural circuits in congenital prosopagnosia. ELife, 6, 1–20. https://doi.org/10.7554/eLife.25069
Song, S., Garrido, L., Nagy, Z., Mohammadi, S., Steel, A., Driver, J., Dolan, R. J., Duchaine, B., & Furl, N. (2015). Local but not long-range microstructural differences of the ventral temporal cortex in developmental prosopagnosia. Neuropsychologia, 78, 195–206. https://doi.org/10.1016/j.neuropsychologia.2015.10.010
Turner, B. O., Paul, E. J., Miller, M. B., & Barbey, A. K. (2018). Small sample sizes reduce the replicability of task-based fMRI studies. Communications Biology, 1(1). https://doi.org/10.1038/s42003-018-0073-z
Witthoft, N., Poltoratski, S., Nguyen, M., Golarai, G., Liberman, A., LaRocque, K., Smith, M., & Grill-Spector, K. (2016). Reduced spatial integration in the ventral visual cortex underlies face recognition deficits in developmental prosopagnosia. BioRxiv, 1–26.
