Reply: Looking beyond indirect lesion network mapping of prosopagnosia: direct measures required

 

We thank Bobes et al.¹ for their interest in our recent work² and commend the authors on applying our results to a new case of prosopagnosia. We also agree that combining our lesion network mapping approach with imaging data from individual patients can be very helpful for understanding the relationship between network structure and function.

In their letter, Bobes et al.¹ study a 46-year-old male (Subject EP) with bilateral posterior cerebral artery strokes and resultant acquired prosopagnosia that avoided the right fusiform face area (FFA) but lined up remarkably well with our proposed ‘prosopagnosia network’ (Fig. 1A). As described in our paper, this network was based on data from 44 patients with acquired prosopagnosia.² Lesion locations from these 44 cases were positively connected to right FFA and negatively connected to left frontal cortex and this was found to be specific to prosopagnosia versus other stroke syndromes. As such, the set of voxels positively connected to right FFA and negatively connected to left frontal cortex defines a brain network (Fig. 1A, purple) that encompasses lesion locations causing prosopagnosia, including our original 44 cases and the new case described by Bobes et al.¹

Figure 1.

A consistent relationship is seen between lesion locations and an identified ‘prosopagnosia network’ with alternative functional connectivity preprocessing algorithms. The traced/segmented bilateral lesion from Subject EP (A, hot colours) as well as the VBM-derived right-sided lesion (A, green) both fall squarely within the proposed ‘prosopagnosia network’ from Cohen et al.² (A, purple), and avoid the right FFA (A, red). Brain regions functionally connected to the VBM-derived right hemisphere lesion were then obtained using the ‘GSP1000’ large resting state functional connectivity database (B), demonstrating strong positive connectivity with the right FFA (B, now shown in blue) as well as negative connectivity with the left anterior prefrontal cortex (APFC), the left middle frontal gyrus (MFG), the dorsal anterior cingulate cortex (ACC), and the left superior frontal gyrus (SFG) (B, red regions). Region–region connectivity between Subject EP’s lesion and the identified critical regions from Cohen et al.² (C, green circles) was also consistent with the original findings from Cohen et al.² when using Global Signal Regression (GSR) (C, left) or when an alternative artefact regression method, such as aCompCor, is used for preprocessing (C, right). Note that in GSR processed data, uninformative connections are distributed around zero, allowing for symmetric thresholds to be used, while in aCompCor processed data, uninformative connections are distributed in the positive range (∼0–0.2), requiring asymmetric thresholds for functional connectivity analyses. (All correlation distributions are significantly different from zero, P < 0.001.) Red lines in box plots indicate medians while stars indicate means.

However, when Bobes et al.¹ examined connectivity with their lesion location using a publicly available normative connectome, they were unable to replicate the finding of negative connectivity between their lesion location and the left frontal cortex. They then used a normative dataset of structural connectivity to explore white matter tracts damaged by the lesion and obtained post-stroke functional neuroimaging data in their patient to explore activation in response to face stimuli and changes in functional connectivity. They conclude that our lesion network mapping results fail to replicate, that functional imaging data from the patient are required, and that modifications to our prosopagnosia network are needed.

We agree that other methods are valuable in understanding and predicting lesion-induced deficits, including functional imaging data from patients.^3,4 However, we were confused by the conclusion that our results failed to replicate. Because the lesion location from Subject EP intersects our ‘prosopagnosia network’, we would predict he would have difficulty with face recognition, which he did. We would also predict that the lesion location would be positively connected to right FFA (which was observed by the authors) and negatively connected to left frontal regions (which was not observed).

To explore this discrepancy in lesion connectivity, we obtained both the original bilateral lesion tracing from the authors (Fig. 1A, hot cluster) and their right hemisphere-only lesion that they used for their lesion network mapping analysis (Fig. 1A, green). We generated a lesion network map for this right hemisphere lesion using the same normative connectome dataset used in our paper (Fig. 1B). We will refer to this as the ‘GSP1000’ normative connectome, which uses global signal regression,⁵ and which we have used extensively in our lesion network mapping studies. We also utilized an alternative connectome, which we will refer to as the ‘GSP100_conn’ normative connectome. This connectome was processed using the CONN toolbox and aCompCor rather than global signal regression.⁶ This connectome has also been used in prior lesion network mapping studies by our group² and uses a similar processing strategy to the connectome used by Bobes et al.¹

Consistent with the 44 cases from our paper, we found that the lesion location from Bobes et al.¹ was positively connected to right FFA and negatively connected to all four regions in the left frontal cortex (Fig. 1B and C). This result was independent of whether we used our original GSP1000 connectome (Fig. 1C, left) or a connectome processed without global signal regression (Fig. 1C, right). In short, we found that the case from Bobes et al.¹ aligns perfectly with our original report.²

The only difference between our analysis and that by Bobes et al.¹ is which normative connectome was used for the analysis, an ongoing issue that can impact lesion network mapping results^4,7 and functional connectivity results more generally.⁵ To facilitate replication and comparison with our prior work, we recently released our preprocessed normative connectome of 1000 typical young adults (age 18–36 years old, M:F 1:1) (https://doi.org/10.7910/DVN/ILXIKS) as well as the code used to generate it (https://github.com/bchcohenlab/BIDS_to_CBIG_fMRI_Preproc2016). As with any preprocessed dataset, this connectome reflects a single set of choices among many researcher degrees of freedom, yet versions of this connectome have yielded significant and meaningful results for over a decade.⁷ We encourage researchers seeking to replicate our lesion network mapping results to utilize this publicly available resource prior to concluding a lack of replication. We also encourage comparison between different connectomes processed in different ways. To our knowledge, there has not yet been a systematic comparison between the Nipype/fMRIprep implementation of the aCompCor algorithm used to process the connectome used in Bobes et al.,¹ and the original implementation of the aCompCor algorithm in the CONN toolbox.⁶ It is possible that these different implementations of aCompCor algorithm yield different results.

Finally, we encourage efforts to complement lesion network mapping with other techniques, including functional imaging in patients and structural disconnection patterns.^3,8,9 Of note, FFA activation in response to faces has been reported in other patients with acquired prosopagnosia¹⁰ and was part of our motivation to look beyond the FFA in our original study.² The structural connectivity and post-injury changes in functional connectivity reported by Bobes et al.¹ are also intriguing, but it is difficult to interpret changes in haemodynamic signals recorded close to a lesion location as well as draw generalizable conclusions from a single subject.¹ At the same time, it is also difficult to assemble a large cohort and acquire functional imaging in patients with relatively rare lesion induced syndromes, which was also part of the motivation for our original study and for lesion network mapping more generally.

Data availability

Data availability is not applicable to this article as no new data were created or analysed in this study.

Competing interests

The authors report no competing interests.