Structural disconnections associated with language impairments in chronic post-stroke aphasia using disconnectome maps

Abstract

Inconsistent findings have been reported about the impact of structural disconnections on language function in post-stroke aphasia. In this study, we investigated patterns of structural disconnections associated with chronic language impairments using disconnectome maps.

Seventy-six individuals with post-stroke aphasia underwent a battery of language assessments and a structural MRI scan. Voxel-based disconnectome-symptom mapping analyses were performed to examine the correlations between disconnectome maps, representing the probability of disconnection at each white matter voxel, and different language scores. To further understand whether significant disconnections were only driven by damaged white matter voxels or also related to seemingly preserved white matter disconnected areas beyond the lesion site, results were compared to voxel-based lesion-symptom mapping analyses using damaged white matter voxels only.

Part of the left white matter perisylvian network was similarly disconnected in 90% of the individuals with aphasia. Surrounding this common left perisylvian disconnectome, specific structural disconnections in the temporo-parietal network were significantly associated with lower performance in syntactic comprehension, syntactic production, and repetition tasks (p < 0.05, after correction for multiple comparisons and adjusting for lesion volume). These language impairments were explained by disconnections in white matter areas that overlapped with and extended beyond white matter lesion sites significant in VLSM analyses investigating the relationship between the same language impairments and white matter damage only. In contrast, language scores related to more distributed neural representations (i.e. aphasia severity, overall auditory comprehension, and naming) were only weakly correlated with the probability of disconnection in specific left-hemisphere white matter areas.

Structural disconnections play a role in the severity of language impairments at the chronic stage, beyond lesion volume. Leveraging routinely available clinical data, disconnectome mapping furthers our understanding of the neural basis of chronic aphasia and provides additional information on anatomical connectivity constraints that may limit the recovery of some language functions.

1. Introduction

Individuals with aphasia may present with various language impairments following an acquired brain injury such as a stroke. Multiple interdependent neurobiological events occur during and after a stroke, resulting in brain tissue death in gray and white matter (Quillinan et al., 2016) and critical disruption of connections between brain regions responsible for language processing. Over the last decades, neuroimaging studies have provided insight into the pathophysiology of these language deficits, and the findings can be broadly categorized into two sets: 1) the effect of focal brain damage on language function, 2) the influence of network-level disruptions on language behavior (Kiran & Thompson, 2019). Here, we will first review these observations in the context of their methodology, focusing on studies using structural data, and then propose a complementary analysis of structural connectivity disruptions that can further our understanding of language impairments in post-stroke aphasia.

Voxel-based lesion-symptom mapping has revealed the relationship between cortical injury and language impairment (VLSM, Bates et al. 2003). Lesion topography is associated with a range of linguistic abilities, such as speech production and speech comprehension (Borovsky et al., 2007; Henseler et al., 2014; Mirman et al., 2015; Price et al., 2010), and more specifically verbal fluency (Baldo et al., 2006), picture naming (Akinina et al., 2019; Døli et al., 2020; Henseler et al., 2014; Piras & Marangolo, 2010), semantic processing (Henseler et al., 2014; Mirman et al., 2015; Schwartz et al., 2009; Walker et al., 2011), repetition (Døli et al., 2020; J. Fridriksson et al., 2010; Henseler et al., 2014), syntactic processing (den Ouden et al., 2019; Dronkers et al., 2004; Lukic et al., 2020; Magnusdottir et al., 2013; Rogalsky et al., 2017), number and word reading (Døli et al., 2020; Piras & Marangolo, 2009) and spelling (Rapp et al., 2016). Traditionally, VLSM uses t-statistics at each voxel of the brain to determine whether the degree of injury is related to language performance (Bates et al., 2003). Lesion-symptom mapping analyses, therefore, provide information on the role of different brain areas in language performance. Nevertheless, they present inherent limitations. First, lesion coverage is heterogeneous across voxels and restricted to particular vascular territories in most studies (Karnath et al., 2019; Rudrauf et al., 2008) which influences the statistical power at each voxel (Rudrauf et al., 2008). Consequently, the analysis may include a spatial bias toward the center of the vasculature territory affected by the stroke (Karnath et al., 2019; Mah et al., 2014). Second, damage to different regions can cause the same language impairment if these regions belong to the same functional network (Julius Fridriksson et al., 2018; Karnath et al., 2019; Price et al., 2017). For these reasons, the prediction power and the interpretation of lesion-outcome associations are limited (Karnath et al., 2019; Kimberg et al., 2007; Price et al., 2017).

Furthermore, stroke damage can disrupt distant regions’ structure and function by modifying their metabolism (Carrera & Tononi, 2014). Von Monakow has defined this neurobiological phenomenon as diaschisis (von Monakow, 1914). Carrera and Tononi have recently extended this notion to ‘connectomal diaschisis’ to describe remote “changes in the structural and functional connectomes, including disconnections and reorganization of subgraphs” (Carrera and Tononi 2014, p. 2419). Hence, language deficits might arise from seemingly undamaged but disconnected regions involved in language processing (Catani & Mesulam, 2008; Price et al., 2017). Interestingly, structural disconnection measures seem to better predict functional connectivity disruption within and between large-scale networks than region-based or voxel-based damage measures (Griffis et al., 2019). Recent studies examining affected anatomical networks in stroke survivors have provided a richer understanding of the relationship between aphasia and its neural underpinnings.

Following the assumption that network-level analyses provide a more comprehensive interpretation of the relationship between clinical symptoms and physiological disruptions caused by the stroke lesion (Catani & Mesulam, 2008), several studies have examined the impact of infarcts on the structural network connectivity and how white matter disruptions relate to language dysfunction. In this paper, the focus is on methods that measure lesions’ effect on anatomical connections (see Zhang et al., 2021 for a meta-analysis of studies using diffusion metrics to investigate white matter integrity in spared tracts). In post-stroke aphasia studies, researchers have mainly used two types of measurements based on white matter tractography data to investigate structural disconnections: (i) reduction in connection density (i.e., percentage number of fibers connected to a cortical region compared to the homologous cortical area), and (ii) a binary measure of tract discontinuity. In the first approach, Bonilha and colleagues found that reduced fiber density at two left hemisphere cortical regions (i.e., Brodmann areas 45 and 22) was associated with the degree of impairment in specific language tasks, but not with overall aphasia severity (Bonilha et al., 2014). The second approach calculates binary measures of disconnection. It considers a tract to be disconnected “if a lesion either disconnects one part of the tract from another, or completely destroys one end of the tract” (Hope and Price 2016, p.1171). Two studies demonstrated that binary disconnection of the left arcuate fasciculus was associated with deficits in naming (Geller et al., 2019; Hope et al., 2016). However, given that this technique relies on a single value for a whole tract, it may be more sensitive to image processing errors, such as misregistration between the lesion map and the probabilistic white matter atlas needed for this method. An error of measurement at one portion of the tract would lead to the opposite category definition for the whole tract (from disconnected to spared and vice versa) (Geller et al., 2019). Mapping disconnected white matter fibers at the voxel level in the whole brain is one way to overcome this limitation. Specifically, detailing the topological distribution of structural network disruption voxel by voxel in post-stroke aphasia provides information to elucidate clinical-anatomical relationships at the level of the affected connectome, from and beyond the lesion site. It also enables one to identify and trace pathways that may be affected by distal effects of stroke damage not easily measurable with actual MRI techniques (Carrera & Tononi, 2014). The most direct way to identify structural disconnections in the whole brain would be to trace fibers that cross each patient’s brain’s damaged area using diffusion-weighted imaging data and fiber-tracking algorithms (Basser et al., 1994; Mori & Zijl, 2002). However, when white matter tracts are directly damaged, reconstruction of the tract’s remaining portions may not be possible (Auriat et al., 2015). Here, we apply an alternative and complementary approach to examine the anatomical substrates of neural disruption in aphasia by constructing disconnectome maps (Foulon et al., 2018). By using a large reference set of high-quality tractograms from healthy controls, these disconnectome maps provide the probability of structural disconnection at each voxel without the need to acquire diffusion-weighted images (Foulon et al., 2018). Specifically, in this work, the probability of disconnection refers to the voxel-wise probability of tracking a white matter fiber in healthy controls. These fibers are then considered disconnected if they also enter the infarcted area when overlaid with the patient’s lesion map. This tool was first used to identify potentially disconnected tracts related to deficits in language processing, decision making, and memory in three well-studied historical patients (Thiebaut de Schotten et al., 2015). Subsequently, this technique has been used to identify disconnected networks following an acquired brain injury related to specific language impairments, such as poor fluency performance (Foulon et al., 2018) and repetitive verbal behaviors (Mandonnet et al., 2019; Torres-Prioris et al., 2019). However, these studies examined only one language component (e.g., speech fluency or repetitive verbal behaviors) per study (Foulon et al., 2018; Mandonnet et al., 2019; Torres-Prioris et al., 2019).

In the present study, we implement a comprehensive clinical-neuroanatomical investigation of the impact of white matter disconnections on a range of language abilities in a large cohort of patients with different types of chronic post-stroke aphasia. We then compare the results from these voxel-based disconnectome-symptom mapping (VDSM) analyses with more standard VLSM analyses limited to damaged white matter voxels. We hypothesize that overall aphasia severity and specific language impairments, such as naming, repetition, syntactic processing, and auditory comprehension, will be associated with disconnections in the left perisylvian connectome due to long-range fiber pathways affected by the lesions. Further, these language impairments will be explained by disconnections that overlap and extend beyond white matter lesion sites associated with each impairment.

2. Materials and methods

2.1. Patients

81 participants with a single left-hemisphere ischaemic stroke who were at least six months post-stroke at time of participation were recruited from three research sites (Boston University, Johns Hopkins University and Northwestern University) between 2015 and 2018 as part of a large-scale study of the Center for the Neurobiology of Language Recovery (http://cnlr.northwestern.edu/). All participants were native English speakers, at least high school educated and had normal or corrected-to-normal vision and hearing. Demographics and neurological history were obtained from medical records and study-specific questionnaires. Exclusion criteria included contraindication for MRI, history of neurological disorder other than a stroke, history of multiple infarcts, history of drug or alcohol abuse and articulatory disorders (apraxia of speech or dysarthria). Five participants out of the 81 have been excluded due to poor imaging data (n=1), different acquisition parameters (n=2), and withdrawal during testing (n=2). A total of 76 individuals with chronic aphasia have been included in the analyses. Participants provided informed consent according to the Declaration of Helsinki. Additional information was provided when needed to ensure participants’ understanding of study protocol before obtaining their written consent. The study was approved by the Institutional Review Boards at all three universities.

2.2. Language assessment

An extensive battery of language assessments was administered to the participants by speech-language pathologists or trained research assistants: the Western Aphasia Battery-Revised (WAB-R) (Kertesz, 2007), Northwestern Naming Battery (NNB) (Thompson et al., 2012) and Northwestern Assessment of Verbs and Sentences (NAVS) (Cho-Reyes & Thompson, 2012). Only specific subtests have been included in the analyses of this study in order to investigate aphasia severity, repetition, naming, auditory comprehension, syntactic comprehension and syntactic production.

2.3. Imaging data acquisition and processing

Imaging data were collected from four different 3 Tesla MRI scanners: a Siemens TIM Trio with a 32-channel head coil and a Siemens Prisma with a 64-channel head/neck coil at Northwestern University, a Siemens TIM Trio with a 20-channel head/neck coil at the Athinoula A. Martinos Center for Boston University, and a Philips Intera with a 32-channel head coil at Johns Hopkins University. Imaging parameters were consistent across the three sites, with the cross-site harmonization verified by the neuroimaging team. High resolution, T1-weighted 3D sagittal volumes were acquired using an MPRAGE sequence (parameters: T1/TE¹/TR = 900/2.98/2300ms, FOV=256×256mm, voxel resolution=1×1×1mm³, 176 sagittal slices, phase encoding direction=A/P).

2.3.1. Lesion mapping

The lesions were traced using a semi-automated procedure on T1-weighted images. First, for each participant, multiple lesion masks were generated by the quality assurance anatomical pipeline available within the Northwestern University Neuroimaging Data Archive (Alpert et al., 2016) through the application of a deep convolutional neural network approach (Wang et al., 2016). This approach uses information about 3D structural image intensity in surrounding voxels as well as contralateral (i.e., right hemisphere) voxels to classify each voxel as belonging to normal or pathological tissue. Second, for each participant, two trained members of the research team independently chose the best lesion mask, and a third member helped resolve any disagreement. When necessary, lesion masks were manually edited in native space using MRIcron software (Rorden & Brett, 2000) (http://www.mccauslandcenter.sc.edu/mricro/mricron/). Spatial normalization of the lesion maps to the MNI space, filtering and resampling to 1×1×1 mm voxels were performed using the 3dQwarp function of the AFNI software through a pipeline in the Northwestern University Neuroimaging Data Archive. Specifically, each pseudo-T1 image with the lesion removed was nonlinearly warped to the MNI template (Brett et al., 2001). The warp field was then applied to the lesion mask to generate the lesion mask in template space. The lesion maps of three participants² were manually drawn using MRIcron software and normalized to the MNI space using SPM 12 instead of this pipeline, due to technical issues. The volume of each lesion map was calculated using in-house Matlab scripts.

2.3.2. Disconnectome mapping

A probability map of white matter tracts’ disconnection was computed for each participant with the “Disconnectome map” tool of the BCBToolkit (Foulon et al., 2018) using diffusion-weighted imaging datasets of 178 healthy controls from the Human Connectome Project database (Thiebaut de Schotten et al., 2020). For each individual with aphasia, the lesion map in MNI space was used as a seed to track all healthy controls’ fibers passing through the lesion in TrackVis (Wang & Wedeen, 2007). Tractographies from the lesion were transformed in visitation maps which represent the number of trajectories that pass through each white matter voxel. A percentage overlap map was then produced by summing, at each voxel in MNI space, the normalized and binarized visitation map of each control subject (Fig. 1). Hence, in the resulting disconnectome map, the value in each voxel takes into account the interindividual variability of tract reconstruction across controls and indicates the probability of disconnection from 0 to 100% for a given lesion (Thiebaut de Schotten et al., 2015). In other words, “probability of disconnection” refers to the voxel-wise probability of tracking a healthy control fiber that enters the infarcted area when overlayed with the lesion map of a patient. Thus, each disconnectome map includes both the part of the fibers that overlaps with the infarcted area (i.e. damaged) and the part that extends beyond the infarcted area (i.e. seemingly spared). A threshold of 50% was applied for this study to ensure generalizability of the results (i.e., at each voxel labelled as disconnected, at least 50% of healthy controls fibers passed through the lesion). After thresholding the disconnectome maps, values at each voxel were either zero (no disconnection) or within the range of 50 to 100% probability of disconnection.

Fig. 1. Voxel-based Disconnectome-Symptom Mapping (VDSM) and Voxel-based Lesion-Symptom Mapping (VLSM) procedures.

For each patient, a disconnectome map was created by overlapping the lesion map of the individual with post-stroke aphasia with healthy control tractograms from the HCP diffusion-weighted images dataset using the Disconnectome map tool of the BCBToolkit. The probability of disconnection at each voxel (e.g. 85% at the voxel highlighted) was then entered into the VDSM analysis. As a subsequent investigation, the intersection between the disconnectome and the lesion map for each patient was computed to only select the damaged white matter voxels within the disconnectome map. A binary value of presence (1) or absence (0) of damage at each white matter voxel was entered into the VLSM analysis.

2.4. Statistical analysis

2.4.1. Voxel-based disconnectome-symptom mapping

The goal of this study was to investigate whether white matter disconnections were related to aphasia severity and other specific language deficits. To this end, we performed VDSM analyses as follows (see more details below): i) thresholding disconnectome maps to exclude voxels with insufficient variance across participants, ii) creating a separate general linear model for each language score, iii) assessing performance of each model with thresholded t-statistic maps and converting them to effect size maps, and iv) identifying white matter tracts and cortical regions involved in thresholded effect size maps.

The power to detect a true relationship between language scores and disconnection probability for a voxel depends on the proportion of patients with and without disconnection in that voxel (Kimberg et al., 2007). Therefore, in accordance with previous VLSM literature (Binder et al., 2016; Borovsky et al., 2007; den Ouden et al., 2019; Dronkers et al., 2004; Klingbeil et al., 2020; Lukic et al., 2020), upper and lower thresholds were set on the disconnectome maps to ensure sufficient variance of the sample at each voxel. Only voxels disconnected in 10% to 90% of the patients (n = 8–68) were included. Clinical-neuroanatomical correlations were performed between individuals’ test scores and the probability of disconnection at each voxel. To this end, the following language scores were included as variables of interest in separate general linear models: Aphasia Quotient (AQ), auditory verbal comprehension and repetition scores from the WAB-R; Confrontation Picture Naming (Nouns) total score from the NNB; Sentence Comprehension Test (SCT) and Sentence Production Priming Test (SPPT) total scores from the NAVS. In total, six general linear models were assessed. Several factors may influence language function in aphasia, such as lesion volume (Døli et al., 2020; Hope et al., 2013; Kertesz et al., 1979; Naeser et al., 1998; Plowman et al., 2011; Watila & Balarabe, 2015), age (Ellis & Urban, 2016; Wallentin, 2018), time post-stroke onset (Holland et al., 2017; Naeser et al., 1998; Pedersen et al., 2004) and handedness (Knecht et al., 2000). Therefore, these variables were included as covariates of no interest in each general linear model. Study center was also added as a covariate of no interest to account for potential differences across testing facilities.

Next, uncorrected t-statistic maps obtained from each general linear model were converted into effect size maps corresponding to Pearson’s correlation coefficient R values between each language score and the probability of disconnection at each voxel. Effect size maps were then thresholded at R>0.3 to display moderate to large effect sizes (Cohen, 2013). In addition to computing effect size maps, we also examined t-statistic maps after multiple comparison correction. Family-wise error rate (FWER) was controlled through permutations (5,000 permutations) following procedure by Winkler and colleagues (Winkler et al., 2014). Specifically, two permutation methods were used to allow more transparency in the results interpretation. First, voxel-wise permutation-based FWER correction, a stringent approach, was used to control for the probability of having one false positive voxel. Second, continuous permutation-based FWER was used as an extension of the standard voxel-wise permutation-based FWER because VDSM results are not expected to be interpreted at the single voxel level. This method allows the transparent reporting of the strength of the evidence within a less stringent framework by quantifying the rate of multi-voxel false positives (Mirman et al., 2018). Thus, results are reported at two rates (v) of false-positive voxels: v=1 and v=100. Finally, Bonferroni correction was applied to voxel-wise permutation-based FWER corrected (v=1) t-statistic maps to control for the number of language scores tested. Results were considered significant below an alpha level of 0.05 (i.e. less than 5% of the permutations had v voxels that exceeded the t-value). Analyses reported above including effect size maps and voxel-wise permutation-based FWER corrected statistical maps (v=1) were conducted with the function randomise from the software package FSL, version 6.0.2 (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/Randomise). Further analyses using continuous permutation-based FWER correction (v=100) were conducted in R (R Core Team, 2018) using the function lsm_regresfast in the package lesymap (Pustina et al., 2018). Figures were produced using MRIcroGL (http://www.cabiatl.com/mricrogl/).

2.4.2. Voxel-based lesion-symptom mapping

Disconnectome maps include disconnected white matter voxels that can be either directly damaged (i.e. direct overlap with the lesion map) or seemingly spared (i.e. part of the fibers non damaged but connected to the lesioned area). To understand whether the significant relationships between disconnections and language deficits were only driven by the lesioned white matter voxels or by disconnected voxels from and beyond the lesioned white matter areas, results that emerged from VDSM were compared to classical voxel-based lesion-symptom mapping analyses limited to the damaged white matter areas. Specifically, the VLSM white matter lesion maps were created by computing the intersection (i.e. overlap) between the overall lesion maps and the disconnectome maps to only obtain damaged white matter voxels within the disconnectome maps in FSL, version 6.0.2 (Jenkinson et al., 2012). Similar to the VDSM analyses, only voxels lesioned in at least 10% of individuals (n>8) were included in order to ensure a sufficient number of participants with a lesion at each voxel. However, the second upper thresholding was not necessary because the maximum lesion overlap (76% of participants) was below 90%. Following the same method as in the VDSM analyses, linear regressions were performed to assess the relationship between the integrity of white matter voxels (lesioned versus non-lesioned) and each language score separately. Lesion volume, age, months post-stroke onset, handedness and site were also included as covariates of no interest in these six analyses. The same multiple comparison correction techniques as in VDSM analyses were used. For comparison purposes, the thresholded effect size maps from the two methods (VDSM and VLSM) were binarized and overlaid for each language test (Fig. 5).

Fig. 5. Overlay and overlap of binarized statistical maps from the VDSM and VLSM analyses.

VDSM = Voxel-based disconnectome-symptom mapping (red, white matter disconnections associated with language deficits) and VLSM = voxel-based lesion-symptom mapping analyses (dark blue, white matter damage associated with language deficits). The overlap between VDSM and VLSM is in light blue. A) Overlays of effect size maps binarized and thresholded at R> 0.3 B) Overlays of binarized t-statistic maps thresholded at p<.05 after correction for multiple comparisons accounting for a probability of up to 100 false positive voxels, and after adjusting for lesion volume. C) Overlays of binarized t-statistic maps thresholded at p<.05 after correction for multiple comparisons accounting for a probability of a single false positive voxel, after adjusting for lesion volume, and after Bonferroni correction across language tests.

2.4.3. Labelling of white matter pathways

We further characterized which white matter pathways were involved in the thresholded effect size maps (R>0.3) of both VDSM and VLSM analyses by overlaying them onto the PANDORA probabilistic atlas of white matter pathways (Hansen et al., 2020). In this study, we used the atlas compiled from the Baltimore Longitudinal Study of Aging tractography data (n=963) extracted via the Automated Fiber-tract Quantification technique. This atlas was selected for the close proximity in age of its population (mean age = 66.2) with the population of individuals with aphasia included in the present study (mean age = 58.3). Second, the Automated Fiber-tract Quantification technique was chosen because its 20 reconstructed tracts include tracts of interest for our study and thought to be involved in language processing based on previous literature. The resulting probabilistic white matter atlas was thresholded at 25%, binarized and overlaid onto the statistical maps using FSL, version 6.0.2 (Jenkinson et al., 2012). The 25% probability threshold ensures a large enough sample size to generalize the atlas labels to our sample while limiting the inter-individual variability that can occur during fiber tracing. To measure the proportion of white matter pathways overlapping with the thresholded statistical maps for each language score, we first computed the volume of the intersection between the atlas mask of each fiber pathway and the thresholded effect size map (R>0.3), and then divided each ‘intersection’ volume by the total volume of the corresponding fiber pathway. These steps were also performed in FSL, version 6.0.2 (Jenkinson et al., 2012). In addition, the Automated Anatomical Labeling atlas in MRIcroGL (https://www.nitrc.org/projects/mricrogl/) was used to identify gray matter regions located at the end of disconnected fibers significantly associated with language scores.

3. Results

3.1. Behavioral results

Participants’ demographics and behavioral scores are available in the Supplementary Table 1. Results are presented for the 76 individuals with chronic aphasia who completed the full battery of language assessments and had good quality imaging data (52 males/24 females; mean age = 58.3, sd = 11.6; mean time post-stroke-onset = 65.0 months, sd = 68.5, range = 8 – 467). Among them, one participant had missing data for NAVS – SPPT and NNB – Confrontation Naming scores but was included for the other analyses.

3.2. Structural maps

Fig. 2 displays the overlay of the 76 lesion maps showing a large coverage of the left hemisphere. The maximum lesion coverage included left periventricular white matter pathways, and left perisylvian white matter (AF) and cortical areas (insula, rolandic operculum). Fig. 3 displays the overlay of disconnectome maps for 76 individuals. The major white matter pathways disconnected in more than 90% of the participants were the left arcuate fasciculus (AF, 23% of the total volume of the tract per the PANDORA atlas), superior longitudinal fasciculus (SLF, 14%), inferior fronto-occipital fasciculus (IFOF, 20%) and posterior part of the inferior longitudinal fasciculus (ILF, 10%) (Fig. 3 C). Given that parts of the AF, SLF, IFOF and ILF were disconnected in almost all individuals with aphasia included in this study, voxels that had greater than 90% of subjects with disconnections were excluded from the VDSM analyses to ensure sufficient variance in the sample analyzed at each voxel, as described in the methods. These excluded voxels are, nonetheless, displayed in effect size maps in Fig. 4 with a dark shade to distinguish these regions disconnected in almost all the participants from the remaining disconnected white matter areas that were included in the analyses and correlated with language scores.

Fig. 2. Lesion overlay.

Plot representing the distribution of lesions in all patients. Slice numbers represent z coordinates of the MNI 152 brain template. Color shades illustrate the increasing number of patients with overlapping lesions (range: 8–58, from cold to warm colors).

Fig. 3. Disconnectome overlay.

Color shades illustrate the increasing number of patients with overlapping disconnectomes (range: 8–76, from cold to warm colors). A) Plot representing the distribution of disconnections for all patients. Slice numbers represent z coordinates of the MNI 152 brain template. B) Sagittal view of the disconnectome overlay. C) Overlap of disconnections that occurred in more than 90% of the participants. IFOF = inferior fronto-occipital fasciculus, AF = arcuate fasciculus, SLF = superior longitudinal fasciculus.

Fig. 4. Structural maps of disconnections associated with language impairments in individuals with chronic aphasia.

Results of the VDSM analyses investigating the relationship between language impairments and the voxel-wise probability of disconnection. A) Effect size maps thresholded at R>0.3 to display moderate to large effects (color shades from red to yellow). Dark color shade represents the overlay of disconnections occurring in more than 90% of the patients and not included in the analyses. B) T-statistic maps thresholded at p < .05, corrected for multiple comparisons at a false positive rate v =100, and after adjusting for lesion volume. C) T-statistic maps thresholded at p < .05, corrected for multiple comparisons at a false positive rate v =1, after adjusting for lesion volume and after Bonferroni correction across language tests. Language scores include aphasia severity (WAB-R Aphasia Quotient), auditory comprehension (WAB-R auditory verbal comprehension), syntactic comprehension (NAVS-SCT) and production (NAVS-SPPT), repetition (WAB-R repetition) and naming (NNB naming). AF = arcuate fasciculus, SLF = superior longitudinal fasciculus, UF = uncinate fasciculus

3.3. Relationships between disconnected tracts and language impairments

In VDSM analyses, all language scores were negatively correlated with voxel-wise probability of disconnection in large left perisylvian and corticobulbar networks at a moderate level (R>0.3) after accounting for lesion volume. The higher the probability of disconnection in these white matter pathways, the lower the language performance across linguistic skills. Notably, the correlations between voxel-wise probability of disconnection and syntactic comprehension (NAVS SCT) and syntactic production (NAVS SPPT) scores demonstrated a large effect size (R>0.5) in left temporo-parietal regions (Fig. 4A). All disconnectome-symptom relationships survived multiple comparisons correction across voxels, when controlling for the probability of up to 100 false-positive voxels, at a significance level of p<.05, after accounting for lesion volume (Fig. 4B). In this condition, the disconnected left white matter areas associated with language scores included white matter tracts underlying the following gray matter regions: (i) the rolandic operculum and insula for aphasia severity and confrontation naming, (ii) the inferior frontal gyrus (pars triangularis and opercularis), rolandic operculum, heschl and superior temporal gyri and the anterior temporal lobe for auditory comprehension, (iii) the rolandic operculum, insula, supramarginal, superior and middle temporal gyri for repetition and (iv) the rolandic operculum, insula, supramarginal, superior, middle and inferior temporal gyri for syntactic comprehension and production, as well as the angular gyrus for syntactic comprehension only. Additionally, patterns of disconnections significantly related to aphasia severity, auditory comprehension and syntactic comprehension and production also included a few voxels in the left corticospinal/corticobulbar pathway.

Finally, when applying the most conservative multiple comparison correction approach to threshold the statistical maps (p <.05, FWER v = 1 and Bonferroni correction) only repetition, syntactic production and syntactic comprehension scores were significantly related to the probability of disconnection in white matter areas underlying the left temporo-parietal junction, extending to the middle temporal gyrus for syntactic comprehension and production (Fig. 4C).

In order to facilitate the identification of networks involved in the effect size maps, the proportions of white matter pathways for which the probability of disconnection (VDSM) or the damage status (VLSM) was moderately negatively correlated with language scores are presented in Table 1. When specifically comparing proportions between VDSM and VLSM, there is consistency in the left hemisphere tracts involved (AF, IFOF, ILF, SLF, thalamic radiations, corticospinal tract and uncinate fasciculus) but the white matter regions involved in VLSM results appear to be more circumscribed. Two exceptions to this statement were the left arcuate and superior longitudinal fasciculi for which the correlations with aphasia severity, naming and auditory comprehension involved lower proportions of the pathways for VDSM than for VLSM analyses. This difference can be explained by the exclusion of voxels in these tracts from VDSM analyses due to insufficient variance in the data compared to VLSM (voxels disconnected in more than 90% of the participants).

Table 1. Percentage of white matter pathways of a probabilistic atlas overlapping with effect size maps of VDSM and VLSM analyses.

Effect size maps (thresholded at R>0.3) represents moderate to large negative correlations between the probability of disconnection (VDSM) or damage status (VLSM) in white matter voxels and language score. IFOF = inferior fronto-occipital fasciculus, ILF = inferior longitudinal fasciculus, SLF = superior longitudinal fasciculus.

	Aphasia severity		Auditory Comprehension		Repetition		Syntactic comprehension		Syntactic production		Naming
Tract name	VDSM	VLSM	VDSM	VLSM	VDSM	VLSM	VDSM	VLSM	VDSM	VLSM	VDSM	VLSM
Corpus Callosum Forceps Major	0%	0%	0%	0%	0%	0%	4%	0%	4%	0%	0%	0%
Arcuate Fasciculus left	28%	28%	31%	27%	30%	22%	27%	15%	28%	10%	12%	19%
Corticospinal Tract left	28%	7%	27%	6%	11%	1%	13%	2%	15%	1%	12%	3%
IFOF left	13%	8%	24%	8%	10%	5%	15%	4%	15%	4%	4%	4%
ILF left	7%	2%	16%	3%	11%	3%	18%	5%	18%	4%	1%	0%
SLF left	25%	30%	22%	27%	19%	17%	14%	9%	11%	3%	12%	20%
Thalamic Radiation left	20%	5%	24%	5%	4%	0%	9%	1%	7%	0%	5%	1%
Uncinate Fasciculus left	14%	7%	30%	7%	6%	5%	6%	2%	5%	3%	6%	3%

3.4. Comparison between VDSM and VLSM

Since the disconnectome maps used in the aforementioned VDSM analyses encompassed both infarcted white matter tissue and potentially spared white matter areas connected to the lesion, we compared the VDSM to more classical VLSM analyses limited to the damaged white matter voxels. Fig. 5 shows that, when examining white matter damage only (i.e. VLSM), all language deficits were negatively moderately correlated (R>0.3) with restricted white matter clusters surrounding the left lateral sulcus. By contrast, when investigating white matter disconnections (i.e. VDSM), structural pathways negatively moderately correlated (R>0.3) with language deficits not only included the damaged regions significant in the VLSM analyses, but encompassed a more extended network of fiber bundles that extended to the left temporal pole, the inferior lateral temporal regions, the inferior frontal gyrus, the orbitofrontal area, the temporo-parietal junction and the brain stem. When correcting for multiple comparisons across voxels and language tests (both v = 1 and v = 100), disconnected white matter areas significantly related to repetition, syntactic comprehension and syntactic production deficits extended into the temporo-parietal junction while damaged white matter voxels significantly related with these scores were located in focal white matter clusters underlying the left superior temporal gyrus. Also, for syntactic comprehension and production, results from the VDSM analyses extended to the left middle and inferior temporal gyri, which was not the case in standard VLSM analyses. Regarding results that survived correction at v = 100, damaged white matter areas related to auditory comprehension deficits included white matter tracts underlying the left rolandic operculum, insula and superior temporal gyrus, while significant disconnections extended to the left white matter pathways connecting the inferior frontal gyrus and the anterior temporal lobe. Conversely, significant disconnections associated with aphasia severity and confrontation naming that survived correction at v = 100, did not extend beyond the damaged white matter areas associated with these language scores.

Relative to the VLSM results, VDSM results suggest that post-stroke chronic language impairments may be related to the disruption of function not only via focal white matter damage but also via disconnections in language networks from and beyond these damaged areas. Second, VDSM identifies regions that are frequently disconnected but less frequently damaged by the infarct in patients with aphasia. Disconnection of such areas, namely the middle and inferior temporal gyri as well as the anterior temporal lobe, may play a significant role in chronic language deficits.

4. Discussion

In this study, using a large dataset of healthy control tractograms as comparison, structural mapping of white matter disconnections was carried out to understand the extended neuroanatomical and behavioral impact of stroke lesions in a relatively large sample of individuals with chronic aphasia. Prior work in aphasia examining the disconnection paradigm has been chiefly used to describe relationships between the AF and repetition deficits (Catani & Mesulam, 2008). Beyond isolated tracts or cortical regions, our results show that different language impairments observed in chronic aphasia are related to structural disconnections in left cortical perisylvian networks. A typical left perisylvian connectome was identified as disconnected in more than 90% of individuals with aphasia resulting from different stroke lesions. While previous studies have looked at (a) white matter integrity in tracts that were not affected by the lesion, (b) reduction in fiber density in spared cortical regions, or (c) binary measures of disconnection, this is the first study, to our knowledge, to map voxel level disconnections in white matter fibers that are associated with the severity of a range of language impairments in chronic post-stroke aphasia.

4.1. White matter disconnections associated with language impairments

By first overlaying the structural disconnectome maps of all participants with chronic aphasia, a consistent anatomical network was identified as similarly disconnected across more than 90 % of individuals despite heterogeneity in language abilities, lesion location, and lesion volume (Fig. 3C). It included parts of both dorsal and ventral left white matter tracts previously described as involved in language processing (Dick et al., 2014; Dick & Tremblay, 2012; Gierhan, 2013): the AF, SLF, IFOF, and ILF. Surrounding this common disconnectome, results of VDSM analyses showed that specific disconnections along left perisylvian tracts play a role in both speech production and speech comprehension deficits, even after accounting for lesion volume (Fig. 4).

Lower abilities in repetition, syntactic comprehension, and syntactic production were related to a higher probability of disconnection in white matter areas beyond the damaged white matter voxels significantly associated with these language impairments in standard VLSM analyses (Fig. 5). Specifically, repetition scores were primarily affected by disconnection in the left AF, extending similar findings from previous studies using diffusion metrics or lesion load (Berthier et al., 2012; Breier et al., 2008; Dick et al., 2014; J. Fridriksson et al., 2010; Kümmerer et al., 2013). Further, our VLSM analyses examining relationships between syntactic processing and white matter damage only are in line with previous findings showing that both syntactic production and syntactic comprehension are related to focal lesions in the white matter underlying the left superior temporal gyrus (den Ouden et al., 2019; Dick & Tremblay, 2012; Lukic et al., 2020). In addition, VDSM illustrates an extension of this network such that disconnections in the left temporo-parietal network impacted by these lesions may also explain up to 50% of the variance in syntactic processing abilities of individuals with chronic aphasia.

When examining auditory comprehension, disconnectome mapping showed that a higher probability of disconnection of tracts underlying the left inferior frontal gyrus and anterior temporal pole was associated with lower auditory comprehension performance, which may involve semantic processing abilities. These results are consistent with the anatomical components of the controlled semantic cognition framework that includes the anterior temporal pole (cross-modal representational system) as well as frontal and temporo-parietal regions (control network) (Lambon Ralph et al., 2017). Previous evidence from diffusion imaging, electrostimulation, and lesion-mapping studies also demonstrated that ventral neural pathways (IFOF and ILF) that connect these two frontal and temporal regions to posterior parts of the brain play a role in semantic processing (Almairac et al., 2015; Duffau et al., 2014; Ivanova et al., 2016; Xing et al., 2017). In our analyses, the disconnected fronto-temporal pathway significantly associated with auditory comprehension impairments could be distinctly identified as the left UF on a white matter atlas. However, whether and how the UF disruption affects auditory comprehension and semantic processing remains a matter of debate as findings have been inconsistent (Dick et al., 2014) and need further investigation. These results did not survive the most conservative thresholding approach in the present study.

Unlike repetition and syntactic processing (comprehension and production) deficits that had clear relationships with disconnections in left temporo-parietal networks, most of the disconnections related to aphasia severity and naming performance did not survive correction for multiple comparisons. One possible explanation may be that distributed representations of language function are associated with aphasia severity and naming abilities (Baldo et al., 2013; Hope & Price, 2016). Thus, while VDSM revealed some degree of anatomical specificity for the other language impairments investigated in this study, heterogeneous presentations of aphasia severity and naming deficits could involve the disconnection of multiple different white matter pathways, resulting in lower statistical power in any specific area of the brain. Additionally, a large part of the left AF was excluded from VDSM analyses because it was disconnected in almost all participants. Disconnection of this tract may play a role in these language behaviors, but this relationship could not be investigated in the present study because of limited variance in our data. In another study that examined binary disconnections in individuals with aphasia, only 67% of the participants showed a disconnection of the AF, and the authors found a significant relationship between the disconnection of the AF and aphasia severity and naming scores (Geller et al., 2019). Other methods have also shown that the integrity of this tract does play a role in the overall aphasia severity at the tract level. A tractography study from our group on a subset of the same data demonstrated that participants whose left AF could not be delineated had a more severe aphasia (Braun et al., unpublished data, 2021). More work is needed to understand the exact contribution of left fronto-temporal disconnections in language impairments related to particularly distributed representations of language function.

Surprisingly, we also found associations between disconnections along the corticobulbar/corticospinal pathway and language scores. However, these associations did not survive the most conservative significance threshold.

4.2. Network disruptions beyond the lesion itself may explain specific language impairments

As described by Karnath and colleagues, when a cognitive function is distributed over a large number of voxels in the brain, the statistical power of lesion-symptom mapping is limited due to mutual exclusion between patients who will have the same impairment from an injury in different brain regions that are part of the same network (Karnath et al., 2019). We speculate that this ‘partial injury problem’ was overcome to a certain extent by disconnectome mapping, as it includes information from lesions that affect the same pathway in a single unit (i.e. disconnected pathway). Specifically, patients with the same language impairment and stroke damage located at different areas along the same anatomical pathway will show distinct lesion-symptom patterns in VLSM and a similar disconnection-symptom pattern in VDSM. In this study, specific language abilities such as repetition and syntactic processing were associated with focal white matter regions underlying the left superior temporal gyrus when looking only at white matter damage, while also related to disconnections along a more extended left temporo-parietal network. Furthermore, VDSM may reveal brain-behavior relationships in areas not typically detected in VLSM studies due to the biased spatial distribution of stroke lesions. For instance, our results confirmed the involvement of tracts underlying the superior and middle temporal gyri in syntactic processing, as previously described in a VLSM study published as part of the same multi-site project (Lukic et al., 2020), and additionally suggest that syntactic deficits could also be explained by disrupted connections with the inferior temporal gyrus which may correspond to terminations of fibers from the AF (Lin et al., 2020). Two recent studies using multivariate predictive models, including one using disconnectome maps, showed that structural connectivity disruptions predicted language performance at a level as good as lesion models using damage location only as input features (Salvalaggio et al., 2020; Yourganov et al., 2016). Despite a potential similar prediction power, mapping clinical-anatomical associations at the disconnectome level allows us to identify the disrupted networks that may have an impact on specific language skills. Here, we showed that these disconnections may impact language processing even at the chronic stage. In other words, this hodological approach (Catani & Ffytche, 2005) informs us on the topological distribution of potential structural modifications in regions remote but directly linked to the infarcted area that are associated with chronic language deficits after a stroke. For instance, disconnection-symptom associations presented here may indicate dysfunctional neural mechanisms such as diaschisis (Carrera & Tononi, 2014; Fornito et al., 2015). However, this interpretation remains speculative and needs further investigation. Further, the continued relationship between left-hemisphere structural disconnection topology and behavior in the chronic stage may reflect anatomical constraints that could limit brain reorganization and language recovery.

4.3. Methodological considerations

In general, the VDSM technique provides useful information on the neural basis of language deficits with data easily accessible from a routine clinical scan (Karnath et al., 2019). Disconnectome mapping partly accounts for brain-behavior relationships in structural networks unified by function (i.e. ‘partial injury problem’), but not for areas unified by the anatomical pattern of stroke damage (i.e. voxels disconnected in a non-random fashion) (Xu et al., 2018). In this regard, two statistical limitations were present in our study. First, a high overlap in disconnection patterns across participants limited the analysis in parts of the language network. A bigger sample size including stroke survivors without aphasia may help overcome this caveat in future studies. Second, mass-univariate models used in this study test each voxel independently (Mah et al., 2014). Even if the localization bias of these models has been characterized primarily on lesion-behavior relationships, we hypothesize that disconnectome mapping results might be similarly biased due to spatial constraints of lesions that affect the connectome investigated and the high collinearity across adjacent disconnected voxels. While we used permutation testing to control for and reduce the number of false positive results across voxels (Karnath et al., 2019), multivariate approaches using machine learning algorithms offer an alternative to account for all voxels at the same time. However, multivariate methods require a large sample of patients (Sperber et al., 2019), and two studies found that multivariate and mass-univariate methods were similarly susceptible to spatial bias (Sperber et al., 2019; Zhao et al., 2020).

Finally, disconnectome probabilistic maps are derived from the HCP data, which is a large set of high-resolution diffusion imaging data of healthy controls that have been preprocessed into tractograms. Therefore, it is important to note that the probability of disconnection is an estimate for characterizing white matter stroke disruptions as the tracts disconnected in patients are hardly traceable with present tractography techniques. The Human Connectome Project dataset contains a majority of young healthy individuals which could reduce the validity of our study due to a mismatch of anatomical data in older individuals. However, three studies have demonstrated that the shape of tracts in disconnectome maps does not significantly change with age (Foulon et al., 2018; Rojkova et al., 2016; Thiebaut de Schotten et al., 2020). With technical advances, diffusion imaging processing software may become more adapted to brains with lesions in order to trace patients’ disconnected tracts and measure microstructural changes from and beyond the infarcted area.

5. Conclusion

In this innovative study, most of the individuals with chronic aphasia presented consistent disconnections in a left perisylvian structural network including parts of the arcuate, superior longitudinal, inferior fronto-occipital and inferior longitudinal fasciculi. Repetition, syntactic comprehension and syntactic production impairments were strongly and significantly associated with disconnections in the left temporo-parietal network. In contrast, language scores related to more distributed neural representations, such as aphasia severity, naming, and overall auditory comprehension, were only weakly associated with specific disconnections in the left perisylvian network. Mapping disconnectome-symptom associations in chronic aphasia allows us to better understand the neural basis of some persistent language impairments by extending previous results to network-level anatomical disruptions that can hinder language function. These results provide complementary information on anatomical connectivity constraints that may limit neural reorganization and language recovery.