Semantic network activation facilitates oral word reading in chronic aphasia

Abstract

People with aphasia often show partial impairments on a given task. This trial-to-trial variability offers a potential window into understanding how damaged language networks function. We test the hypothesis that successful word reading in participants with phonological system damage reflects semantic system recruitment. Residual semantic and phonological networks were defined with fMRI in 21 stroke participants with phonological damage using semantic- and rhyme-matching tasks. Participants performed an oral word reading task, and activation was compared between correct and incorrect trials within the semantic and phonological networks. The results showed a significant interaction between hemisphere, network activation, and reading success. Activation in the left hemisphere semantic network was higher when participants successfully read words. Residual phonological regions showed no difference in activation between correct and incorrect trials on the word reading task. The results provide evidence that semantic processing supports successful phonological retrieval in participants with phonological impairment.

1. Introduction

Disruption to the phonological system is a common occurrence in ischemic stroke due to its central location in the left middle cerebral artery (MCA) territory (Caviness et al., 2002; Mohr et al., 1992; Pillay et al., 2014). The phonological system supports long-term representation, retrieval, and short-term maintenance in awareness of the sensory forms of words and word-like (e.g., pseudoword) entities, as well as more schematic properties of these structures, such as language-specific “rules” governing serial dependencies and alternations (Anderson, 1985; Levelt, 1993; Sapir, 1921). Phonological impairments are common to several forms of aphasia (Damasio & Damasio, 1980; Dick et al., 2001; Indefrey & Levelt, 2000, 2004), manifesting as phoneme-level paraphasia and omission errors on tasks requiring access to word-sound forms, such as oral reading (Rapcsak et al., 2009; Ripamonti et al., 2014), rhyming (Pillay et al., 2014; Rapcsak et al., 2009), repetition (Baldo et al., 2012; Fridriksson et al., 2010; Rapcsak et al., 2009), and picture naming (Baldo et al., 2013; Kohn & Goodglass, 1985; Schwartz et al., 2012). Converging lesion-symptom mapping, functional neuroimaging, and cortical stimulation studies have localized the phonological system to left perisylvian cortex and underlying white matter pathways, including the posterior superior and middle temporal gyri, supramarginal gyrus, inferior frontal gyrus, and arcuate fasciculus (Booth et al., 2002; Bruno et al., 2008; Buchsbaum et al., 2011; Butler et al., 2014; Indefrey & Levelt, 2004; Lacey et al., 2017; Mirman et al., 2015; Pillay et al., 2014; Roux et al., 2012; Wilson et al., 2011). Damage to the phonological system commonly affects oral reading, and post-stroke alexia is a common deficit (Brookshire et al., 2014).

While damaged language systems undergo varying degrees of neural reorganization during recovery, the precise mechanisms whereby reorganization improves performance are unclear (Fridriksson et al., 2009; Lee et al., 2017; Meinzer et al., 2006). One little-explored avenue for addressing this problem is the analysis of brain activity underlying trial-to-trial variability (Pillay et al., 2018; Postman-Caucheteux et al., 2010). People with aphasia (PWA) often show only partial impairments on a given task, manifested by errors on some but not all trials. This trial-level variation offers a potential window into understanding how damaged language networks function and which behavioral and neural factors are associated with performance. In a previous fMRI study of participants with phonological system damage and preserved semantic function, the left angular gyrus (AG) was activated more when participants read words aloud correctly than when they made errors (Pillay et al., 2018). The AG is involved in word reading (Binder et al., 2005; Boukrina & Graves, 2013; Graves et al., 2010) and semantic processing (Binder et al., 2009) in healthy controls, thus we speculated that these participants recruited their undamaged semantic system to “boost” function in the damaged phonological system on successful word trials.

The semantic system supports long-term storage and retrieval of general factual knowledge, such as knowledge about objects, events and actions, abstract concepts, social behavior, and culture (Murphy, 2002; Tulving, 1972). Activation of this information by familiar word forms underlies word comprehension, and self-initiated activation of this information underlies propositional language production (Levelt, 1993). Converging lesion-symptom correlation and functional neuroimaging studies show that the semantic system is broadly distributed in bihemispheric temporal, parietal, and frontal lobe cortices (Binder et al., 2009; Lambon Ralph et al., 2017). Notably, much of this network lies outside the core MCA territory most commonly damaged in ischemic stroke. Although reading aloud is considered a task with minimal semantic processing demands, some neural network models posit a contribution from semantic codes in computing phonology (Plaut et al., 1996; Woollams et al., 2007), a claim supported by facilitatory effects of semantic variables (e.g., imageability) on oral reading speed in healthy individuals (Balota et al., 2004; Shibahara et al., 2003; Strain et al., 1995; Woollams, 2005). There is also evidence that word familiarity and imageability, two factors that index meaningfulness, facilitate reading in people with aphasia (Coltheart et al., 1980; Dickens et al., 2019; Woollams et al., 2018). Here, we sought to examine whether semantic network activation supports reading in chronic aphasia after damage to the phonological system.

Direct evidence relating activation in the semantic network with better reading performance in PWA has not yet been provided. Although our previous study showed performance-related activation in the left AG, our hypothesis that this activation represents semantic processing is based purely on “reverse inference” from localization studies in healthy people. Such inference is unlikely to be valid when applied to patients with left hemisphere damage. Moreover, non-semantic roles have been attributed to the left AG, including phonological, cognitive control, and attention functions (Carter & Huettel, 2013; Noonan et al., 2013; Pugh et al., 2000).

The current study addresses the semantic boost hypothesis directly using a “functional localizer” approach to identify semantic regions of interest within the same PWA sample whose word reading performance was previously reported. Given the a priori hypothesis that semantic system activation contributes to oral reading success, we used a separate fMRI protocol to identify semantic and phonological regions, then used this functionally defined network as a region of interest to test for facilitatory effects of activation on reading performance. This method minimizes uncertainty in ascribing an underlying semantic role to given brain regions. We primarily focus on semantic regions in the left hemisphere because they likely have stronger connections with left-lateralized phonological networks than do right hemisphere semantic regions, and therefore are more likely to play a role in phonological retrieval (Lambon Ralph et al., 2001). However, given the long-standing debate regarding the role of the right hemisphere in aphasia recovery (Gainotti, 2015; Heiss et al., 1999; Karbe et al., 1998; Saur, 2006), and previous work suggesting correlated task-evoked activations in left and right semantic regions (Griffis et al., 2017), we also explored right hemisphere semantic regions of interest. We hypothesized that activation in the left hemisphere semantic network contributes to successful oral reading performance, which was tested by comparing activation levels in these regions during reading trials.

2. Materials and methods

2.1. 1 Participants

Participants in the study included 21 prospectively recruited individuals with chronic left hemisphere ischemic stroke (10 women, 11 men) who showed impaired phonological retrieval (pseudoword rhyme matching task) and relatively intact semantic retrieval (object similarity matching task) on screening tests used as inclusion criteria. The participants reported in this study (see Table 1 for demographic data) overlap completely with the group described by Pillay, et al. (2018). Scores on the screening tests used to identify individuals for inclusion were transformed to standardized z-scores based on performance data from 24 representative healthy, right-handed, age-and education-matched controls (Pillay et al., 2018). All participants were significantly impaired (z < −1.67) on the phonological retrieval test, which assessed the ability to match written pseudowords on rhyme similarity, but performed within the normal range (z > −1.67) on the semantic retrieval test, which assessed the ability to match object pictures on semantic similarity. Mean z-score on the phonological retrieval test was −5.60 (SD 2.74), whereas mean z-score on the semantic retrieval test was −0.61 (SD 0.76). Additional information on the participants’ behavioral profiles is shown in Supplementary Table 1, using the same healthy control sample mentioned above to compute z-scores. These additional tests, which assessed phonological and semantic processes using semantic word matching (similar to the task given in the scanner, described below), reading, word-to-picture matching, picture naming, pseudoword repetition, and phonological short-term memory, are provided to further characterize the aphasia profile in each patient, but were not used in the context of the present study, and are not discussed further. All participants were at least 180 days post-stroke, native English speakers, and pre-morbidly right-handed according to the Edinburgh Handedness Inventory handedness quotient (M = 84.0, SD = 25.9) (Oldfield, 1971). Lesions included 17 MCA infarcts, 2 combined MCA/anterior cerebral artery infarcts, and 2 combined MCA/posterior cerebral artery infarcts. The study was approved by the Medical College of Wisconsin Institutional Review Board and undertaken with informed consent from participants in accord with the Declaration of Helsinki.

Table 1.

Aphasia participant demographic and lesion volume data.

Variable	Mean	SD	Range
Age (years)	56.4	12.5	30 – 80
Education (years)	14.7	3.2	8 – 20
Days Post Onset	1134	1491	180 – 6732
Lesion Size (ml)	73.4	58.6	6.7 – 227.0

2.2. FMRI localizer tasks

In a single scanning session that included both the Reading protocol described in Pillay et al. (2018) and the Matching protocol described in detail below, participants performed two-alternative forced-choice matching tasks designed to elicit either phonological (Phonological task) or semantic (Semantic task) processing, as well as a nonlinguistic control task (Control task) (see Fig. 1 for examples). Stimulus presentation and response recording were controlled with an E-prime script (Psychology Software Tools; Sharpsburg, PA). In all tasks, a sample stimulus was displayed in the center of the visual display, and participants were given two choice options below, creating a stimulus triad. The three stimuli presented in the Phonological task were all pseudowords, whereas the stimuli in the Semantic task were all concrete nouns. In the Phonological task, participants selected the option that best rhymed with the sample, with the choice options including a plausible rhyme that did not match the sample orthographically (e.g., jole for the sample oal) and a close but implausible rhyme (bule for the sample oal). In the Semantic task, participants selected the option most similar in meaning to the sample from either a close semantic neighbor, defined as having a high degree of feature overlap with the sample (e.g., bib for the sample apron), or a more distant semantic neighbor having fewer features in common with the sample (tuxedo for apron). The Control task required matching unfamiliar (Japanese) character strings on font size. This task provides a baseline for the Phonological task by providing controls for visual sensory, motor, and general attention and executive processes, while minimizing any phonological processing. Given that some of the participants had motor deficits affecting the right hand, all responses were made using a button box operated with the left hand. Accuracy and reaction time (time from stimulus onset to button response) were recorded automatically by the E-prime script. A group of 16 right-handed, age- and education-matched healthy controls (mean age = 57.4, SD = 9.6; mean education = 15.1, SD = 2.1) normed the fMRI stimuli listed above, and patient z-scores compared to this healthy control sample are provided in Supplementary Table 1. A list of stimuli used in the localizer tasks is provided in Supplementary Table 2.

Fig. 1.

Example trials in the fMRI localizer protocol. Left image: Phonological (rhyme matching) task. Middle image: Semantic (meaning matching) task. Right image: Control (size matching) task.

Timing and order of stimulus presentation were controlled with E-prime software and synchronized to MRI data acquisition. Standard 800 × 600 RGB output was sent from a PC to an LCD video projector. Images were back-projected onto a screen placed 240 cm from a head-coil-mounted prism lens just above the participant’s eyes. Individual stimuli were presented in the center of a screen in white font on a black background, with each item in the triad typically subtending a maximum horizontal visual angle of 2.0–2.5°. Each test triad was displayed for 4000 ms. A mixed block and event-related design was used, with individual trials within blocks separated by a jittered inter-stimulus interval of 2–4 s during which a fixation cross was displayed. This allowed for event-related analysis to be performed separately on correct and incorrect trials. There were four localizer runs, each approximately 8 min long, each containing 2 blocks of each condition (Semantic, Phonological, Control), with 9 trials per block (18 trials of each condition within each run; 72 unique trials). Each block was approximately 74 s long, with 4 s between blocks. Task specification was indicated by a 1-word instruction cue (Rhyme, Meaning, or Size) presented for 4 s prior to each block. The order of conditions was pseudo-randomized, with no consecutive blocks of the same condition. Participants received scripted step-by-step instructions and practiced each task extensively with feedback until performance exceeded chance prior to entering the scanner.

2.3. FMRI oral reading task

During the same imaging session, participants performed an overt oral reading task, in which 72 nouns were presented over the course of six scanning runs, using the same visual display as in the localizer tasks. Words included 36 regular and 36 irregular words, matched on imageability, length, and frequency (Balota et al., 2004). Further information regarding imageability, length, and frequency are provided elsewhere (Pillay et al., 2018). As a low-level baseline task (different from the Control task used in the matching protocol), 72 strings of unfamiliar (Japanese) characters, matched in length to the words, were randomly mixed with the words. All stimuli were presented for 2000 ms and followed by a fixation cross of variable duration (2 to 26 sec) optimized for event-related analysis. Participants were asked to read each word aloud, and to say “junk” when presented with unfamiliar characters. Spoken responses were captured via a fiber-optic dual-channel noise-canceling microphone (FOMRI-III, Optoacoustics) positioned near the participant’s mouth. Responses were analyzed off-line for accuracy and reaction time (time from word onset to response onset). Word reading responses were considered incorrect if the participant failed to produce a complete response or failed to produce a correct response. Incomplete or incorrect responses that were followed immediately by correct responses were counted as correct as long as the correct response occurred before the next trial.

2.4. MRI acquisition

MRI data were acquired on a 3 T GE Excite whole-body scanner. Four participants were scanned with a 32-channel head coil, and the remaining 17 were scanned with an 8-channel head coil. No covariates were included for the different head coils in analysis. A gradient-echo, echo-planar imaging sequence (flip angle = 77°, echo time = 25 msec, repetition time = 2000 msec, NEX = 1, field of view = 192 cm, matrix = 64 × 64, slice thickness = 3 mm, inter-slice gap = 0.5 mm, 36 axial slices) was used for isotropic (3 × 3 × 3 mm) whole-brain fMRI data acquisition. Head movement was monitored during scanning using real-time image registration (Cox & Jesmanowicz, 1999). High resolution T1-weighted anatomical reference images were obtained using a spoiled gradient-recalled (SPGR) sequence (flip angle = 12°, echo time = MinFull, T1 prep = 450 msec, repetition time = 8.2 msec, NEX = 1, voxels 1.0 × 1.0 × 1.0 mm, 162 axial slices).

2.5. FMRI data analysis

Image processing and statistical analyses were performed using the Analysis of Functional NeuroImages (AFNI) software package (Cox & Hyde, 1997). Preprocessing steps included slice timing correction and affine image registration of each EPI time series. The first four images in each time series were discarded prior to regression analysis to avoid saturation effects. Translation and rotation parameters, estimated during registration, were saved for use as noise covariates. The EPI volumes were then registered to the T1 anatomical image using a modality-specific cost function based on weighted local Pearson coefficients (Saad et al., 2009) using the AFNI script “align_epi_anat.py”. Image time points contaminated by residual artifactual transients were identified using automated routines in AFNI (3dToutcount) and excluded from the analysis. This method resulted in an average exclusion of 3.2 % of the image volumes.

BOLD signal changes were analyzed using a multiple linear regression model in the AFNI program 3dDeconvolve. No smoothing was done prior to the regression analysis. Stimulus regressors were created for each task condition by convolving the stimulus onset with a canonical gamma variate hemodynamic response function. For the localizer scans, the time series regressors coded: 1) Correct Semantic trials, 2) Incorrect Semantic trials, 3) Correct Phonological trials, 4) Incorrect Phonological trials, 5) Correct Control trials, 6) Incorrect Control trials, and 7) Instruction Cue. Six motion vectors computed during image registration were included as covariates of no interest. For the oral reading scans, the regressors coded Correct word reading trials, Incorrect word reading trials, and Baseline (“junk”) trials, with covariates of no interest including the six motion estimates and an additional covariate representing the averaged signal in the ventricles to estimate speech movement artifact (Graves et al., 2007).

2.6. Lesion tracing and template registration

Lesioned areas were identified using a semi-automated procedure to create an individual lesion map for each participant (Pillay et al., 2014). Each participant’s anatomical image and associated lesion map were then morphed to a stereotaxic template (“Colin N27′′) in Talairach standard space (Talairach & Tournoux, 1988) using AFNI’s nonlinear registration routine 3dQwarp. The cost function computed during the morphing process ignored voxels within the lesion map. The parameters used for warping each individual’s anatomy to the template were then applied to the lesion map and functional images. Fig. 2 shows an overlap of the lesions in template space, thresholded at a minimum overlap of 6 participants. The lesions primarily affected perisylvian cortex (superior temporal, supramarginal, pre- and postcentral, and inferior frontal gyri) implicated previously in phonological processing, and insula and underlying white matter.

Fig. 2.

Lesion overlap map, shown on serial sagittal sections through the left hemisphere. Colors indicate the number of participants with a lesion at each voxel location, as coded on the color scale at right. Numbers below each panel indicate stereotaxic × coordinates. Slice locations match those in Fig. 3 for ease of comparison.

2.7. Semantic and phonological ROI analyses

A group-level Semantic network ROI was defined from the localizer maps by contrasting the Semantic and Phonological beta coefficient maps produced by multiple regression for correct trials only. Using the Phonological task as a control for visual, phonological, general executive, and motor responses engaged by the Semantic task produces a relatively specific map of neural activation related to semantic processing. This contrast was first computed at the individual participant level, then a Gaussian kernel of 7 mm FWHM was used to smooth these difference maps prior to stereotaxic transformation. These contrast maps were then analyzed with a single-sample t-test at the group level. After retaining only positive (Semantic > Phonological) values, the resulting map was thresholded at a lenient voxel-wise p <.05 and minimum cluster size of 500 μl. A lenient criterion was adopted to ensure sampling of the majority of the semantic network even at the expense of including some voxels that would not pass a stringent correction for multiple comparisons.

The Phonological network ROI was defined by contrasting the Phonological and Control beta coefficient maps (correct trials only), followed by group-level t-testing and thresholding identical to that used to create the semantic network ROI. This contrast produces a relatively specific map of phonological activation by using the Control task to control for visual sensory, general executive, and motor response processes engaged by the Phonological task. Note that the Semantic task is assumed to incidentally engage phonological processing (Carr et al., 1982; Pugh et al., 1996), making the Semantic task unsuitable as a control for phonological system mapping.

Semantic and Phonological network ROIs were then individualized to each participant by masking out voxels in the group ROI that had negative values in the corresponding participant’s map. For example, only voxels in the Semantic network ROI that also had positive beta coefficient values in an individual’s Semantic - Phonological contrast map were included in that individual’s Semantic network ROI. These methods ensured that voxels included in each individual’s Semantic and Phonological network ROIs were highly likely to be involved in the semantic or phonological process of interest, while also enforcing similar ROI locations across individuals. In addition, the network ROIs were further individualized by excluding voxels that overlapped with the lesion volume.

In line with our central hypothesis that participants with phonological impairment use undamaged portions of the semantic network to support oral word reading, we predicted that the left hemisphere semantic system would be more active during Correct than Incorrect word reading trials. Average voxel beta coefficients within the left hemisphere semantic ROI were computed for each of the word reading conditions (correct and incorrect) relative to the baseline reading task (saying “junk” for unfamiliar character strings) in each participant. To account for variability of gain in MR signal across participants, these values were scaled on an individual basis to the average raw signal value across all brain voxels. An identical analysis was carried out using the right hemisphere semantic ROI.

To assess whether any observed effects related to word reading accuracy were unique to the semantic network, an identical analysis of activation during Correct and Incorrect word reading trials was carried out using the left and right phonological network ROIs. Data from all ROIs were analyzed using repeated-measures ANOVA, with hemisphere (Left vs Right), network (Semantic vs Phonological) and reading condition (Correct vs Incorrect) as within-subject factors, and the number of voxels that were lesioned and removed from each network (Phonological and Semantic Network Damage) included as covariates, as differential network activation may depend on the degree of integrity of the system. This was followed by post-hoc paired t-tests adjusted for multiple comparisons (Bonferroni). We also correlated performance on the fMRI Semantic or Phonological task with Word performance activation in residual semantic and phonological networks to further explore this relationship.

3. Results

3.1. Task performance

Details regarding performance on the oral word reading task were presented previously (Pillay et al., 2018). In brief, accuracy on this task averaged 75.9 % (SD = 20.4), and the number of errors ranged from 2 (2.8 %) to 52 (72.2 %). Most errors were phonologically related to the target item (M = 10.4, SD = 8.2), followed by errors that were unrelated to the target item (M = 3.7, SD = 6.7). Semantically related errors were rare (M = 1.7, SD = 1.4), as were omissions (M = 1.4, SD = 2.2). Performance on regular word trials (M = 79.4, SD = 21.0) was better than performance on irregular word trials (M = 72.5, SD = 21.1) paired t(20) = 4.75, p <.001. Correct and incorrect responses did not significantly differ on reaction time. See Pillay et al., 2018 for more detailed description of reading performance.

Performances on the localizer tasks are summarized in Table 2. As expected, participants performed better on the Semantic task than on the Phonological task (paired t(20) = 6.03, 2-tailed p <.001). Performance on the Control task was also better than the Phonological task (paired t (20) = −4.66, 2-tailed p <.001). Semantic and Control task performances did not differ (paired t(20) = 0.62, 2-tailed p =.52). Semantic and Phonological scores were correlated with each other (r = 0.68, p <.001), as well as with the Control task (r = 0.49, p =.024 and r = 0.54, p =.012, respectively). Semantic and Phonological scores were both negatively correlated with lesion volume (r = −0.55, p =.010 and r = −0.44, p =.048, respectively). None of the scores were correlated with age, education, gender, or days post-onset. Performance on all three localizer tasks were correlated with word reading accuracy, with Phonological scores most strongly correlated (r = 0.75, p <.001), followed by Semantic scores (r = 0.60, p =.004), and finally Control scores (r = 0.48, p =.029). Participants responded faster during the Control task compared to both Semantic and Phonological tasks (both p <.001). There was no difference in reaction time between Semantic and Phonological tasks (p=.419).

Table 2.

Participant fMRI localizer task performance data.

Aphasia Group
Variable	Mean	SD	Range
Semantic Task accuracy	80.1	11.3	41.7 – 94.4
Phonological Task accuracy	62.2	18.4	27.8 – 91.7
Control Task accuracy	78.3	13.0	52.8 – 95.8
Semantic Task RT (ms)	2783	230	2417 – 3184
Phonological Task RT (ms)	2865	459	1530 – 3648
Control Task RT (ms)	2394	508	1642 – 3543

3.2. FMRI localizer and ROI analysis

Areas activated by the Semantic task relative to the Phonological task (Semantic network), shown in red in Fig. 3, included the bilateral angular gyrus, posterior cingulate gyrus, and dorsomedial prefrontal cortex; left inferior frontal, orbital frontal, and middle temporal gyri; and right parietal operculum. The Phonological network (phonological task > control task), shown in blue in Fig. 3, included left lateral prefrontal and premotor cortex, right inferior frontal gyrus/insula, bilateral ventral occipitotemporal cortex (lateral fusiform gyrus and occipitotemporal sulcus), and bilateral supplementary motor area. Stereotaxic coordinates of activation peaks within these clusters are listed in Table 3. Whole-brain non-binarized versions of the localizer maps are shown in Supplementary Fig. 1.

Fig. 3.

Regions-of-interest (ROI) networks defined by the fMRI localizers. Numbers indicate stereotaxic × coordinates. The Semantic network (semantic task > phonological task) is shown in red, and the Phonological network (phonological task > control task) is shown in blue. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Table 3.

Semantic and Phonological activation peaks in Talairach space.

Semantic > Phonological	Coordinates (RAI)
	Location of peak	x	y	z	t-score
	R orbital frontal cortex	−2	−60	−10	5. 913
	L orbital frontal cortex	1	−61	−9	5.366
	L angular g.	49	61	20	5.179
	L superior frontal g.	3	−55	29	4.728
	R superior frontal g.	−8	−62	11	4.471
	R angular g.	−45	71	22	4.174
	L posterior cingulate g.	2	44	27	3.940
	R parietal operculum	−55	35	20	3.327
	R posterior cingulate g.	−5	45	30	3.124
	L. middle temporal g.	52	11	−11	3.041
	L inferior frontal g.	49	−26	14	2.952
Phonological > Control	L ventral temporal g.	43	55	−18	6.028
	L inferior frontal g.	39	−14	27	5.563
	R superior frontal g./SMA	−1	−4	57	5.213
	L superior frontal g./SMA	2	−3	50	4.951
	R inferior frontal g.	−47	−10	−1	3.801
	R ventral temporal g.	− 26	65	−15	3.107

Repeated-measures ANOVA of the BOLD responses in individualized Left and Right Hemisphere Semantic and Phonological networks during Correct and Incorrect word reading revealed a significant crossover interaction between hemisphere, network, and word reading accuracy (F(1, 18) = 10.41, p =.005, η² = 0.37). There were no main effects of reading accuracy (p =.67), hemisphere (p =.12), or network (p =.49). Two-way interactions between hemisphere and network (F(1, 18) = 2.21, p =.15, η² = 0.11), reading accuracy and hemisphere (F(1, 18) = 0.78, p =.39, η² = 0.04), and reading accuracy and network (F(1, 18) = 1.07, p =.32, η² = 0.06) were also not significant. Planned contrasts revealed that within the Semantic network, activation was significantly greater for Correct than Incorrect word reading trials only in the left hemisphere (paired t(20) = 2.59, 2-tailed p = 0.018), but not the right hemisphere (paired t(20 = 0.79, 2-tailed p =.441). There was no significant difference between Correct and Incorrect word reading trials within the left (paired t(20) = 0.40, 2-tailed p = 0.697) or right (paired t (20) = 1.42, 2-tailed p = 0.172) Phonological network. Mean BOLD signal changes during Correct and Incorrect oral word reading in Semantic and Phonological network ROIs are shown in Fig. 4.

Fig. 4.

Mean BOLD signal change during Correct and Incorrect oral word reading (in arbitrary units relative to a nonsense character string baseline) in Semantic and Phonological network ROIs. Error bars indicate standard error of the mean. * = p <.05.

The Phonological Network Damage (mean amount of network damage = 16.9 %, SD = 15.2 %, range 0.4–51.2 %) covariate was significantly related to the interaction between reading accuracy, hemisphere, and network (F(1, 18) = 5.05, p =.037, η² = 0.22). Greater damage to the Phonological Network was associated with lower BOLD signal during Incorrect word trials in the left Semantic ROI (t(18) = −2.66, p =.016, η² = 0.28; estimated coefficient of −3.7e⁻⁶, 95 %CI [−6.64e⁻⁶, −7.83e⁻⁷]) and in the right Semantic ROI (t(18) = −3.24, p =.005, η² = 0.37; estimated coefficient of −5.6e⁻⁶, 95 %CI [−9.17e⁻⁶, −1.95e⁻⁶]). Phonological Network damage was not related to BOLD signal in the left or right Semantic ROIs during Correct word trials, or to BOLD signal in the left or right Phonological ROIs during Incorrect or Correct word trials. There was no significant relationship between the Semantic Network Damage (mean amount of network damage = 13.1 %, SD = 10 %, range 0.5–39 %) covariate and the 3-way interaction (F(1, 18) = 0.28, p =.869).

One participant (4572) performed at chance on both the Phonological and Semantic localizer tasks, which is likely to have contributed some noise to the analyses. Removal of this patient from all analyses did not significantly change the results of the main 3-way interaction between hemisphere, network, and word reading accuracy (F(1,17) = 9.88, p =.006).

4. Discussion

Many gaps remain in our neurobiological understanding of how language networks reorganize in response to focal damage, and which alterations in particular facilitate recovery. In the current study we exploited the fact that most individuals in the chronic recovery state show incomplete deficits, allowing group and item-level analysis of brain activity as a function of performance success when reading. People with impaired oral reading due to phonological system damage sometimes show strong facilitatory effects of word familiarity, concreteness, and imageability, a pattern interpreted as evidence that activation of lexical or semantic information can facilitate phonological retrieval (Coltheart et al., 1980; Friedman, 1995; Patterson, 1996; Tree, 2008). Here we show for the first time that successful word reading in people with phonological system damage is specifically associated with increased neural activation in the (functionally defined) semantic system. The results have practical implications for the optimal design of future rehabilitation and noninvasive electrical stimulation approaches for promoting recovery in people with aphasia.

In our previous study on this topic, we approached the question empirically by searching for brain regions where activation was stronger for words read correctly than words read incorrectly (Pillay et al., 2018). One limitation of this approach is that the need for a strict threshold to correct for multiple comparisons across the brain could place undue focus on “tips of the iceberg” in a highly distributed network. The semantic network is known to be widely distributed (Binder et al., 2009), thus it is possible that the semantic contribution to oral word reading is also highly distributed, and that the facilitatory contribution of any one voxel in this network is relatively small. Another limitation is uncertainty regarding any claims on the type of process activated in regions showing a word performance effect. Although the left angular gyrus region identified in the prior study overlaps entirely with the angular gyrus cluster identified in a previous meta-analysis of semantic imaging studies (Binder et al., 2009), there remains the possibility that such “reverse inference” is invalid when applied to people with aphasia with variable amounts of left hemisphere brain damage.

The current study addresses both of these limitations. Given the a priori hypothesis that semantic system activation contributes to oral reading success, we used a separate fMRI protocol to identify semantic regions in the same participant cohort, then used this functionally defined network as a region of interest to test for facilitatory effects of activation on word reading performance. The results provide strong confirmation of the main hypothesis and also suggest that the facilitative activation in the semantic network is likely to be more distributed than was suggested by the previous analysis.

The results of the localizer study show a left hemisphere semantic network very similar to the one reported in functional imaging studies of healthy participants when adequate controls are included for word form (orthographic and phonologic) and domain-general cognitive control (attention, working memory) processes (Binder et al., 2009). This similarity is not surprising given that participants in the current study were required to have relatively intact semantic performance on a screening measure, and that many of these brain regions, including angular, dorsomedial prefrontal, orbital frontal, and posterior cingulate cortex, lie outside the central perisylvian MCA territory damaged in our participants. Level of activation in this semantic network appears to be the main factor determining whether word reading is successful on a given trial. This benefit was most apparent in the left hemisphere semantic network recruitment, and suggests that for PWA with phonological retrieval impairments, semantic activation may act as a ‘compensatory mediator’ for word reading ability. The right hemisphere semantic network showed a similar pattern of relative activation (less deactivation compared to baseline) for Correct compared to Incorrect trials, however this numerical difference did not approach significance. Activation associated with incorrect word reading within both the left and right semantic network was related to the degree of phonological network damage, suggesting that the extent of damage to the phonological network modulates the degree to which word reading depends on semantic network activation.

The phonological network in our participants, defined here by a contrast between a visual rhyme matching task and a size matching task with unpronounceable character strings, likely differs from the healthy phonological network. Key components of the healthy phonological network lie in the posterior perisylvian (supramarginal and superior temporal) cortex (Buchsbaum et al., 2011; Mirman et al., 2015), and damage to and connections with this region almost certainly accounts for the phonological impairments in our sample (Pillay et al., 2014). Residual left hemisphere phonological processing in our participants appears to rely mainly on lateral and inferior frontal and ventral occipitotemporal cortex. The left inferior frontal and precentral gyri have often been implicated in aspects of phonological processing, including syllabification, maintenance of phonological information in short-term memory, and articulatory planning (Indefrey & Levelt, 2004; Tourville & Guenther, 2011; Wager & Smith, 2003). It is somewhat unlikely that phonologic representations themselves are localized in the precentral and frontal cortices (Buchsbaum, Hickok, & Humphries, 2001; Price, 1998; 2010). Rather, these regions have been linked with response selection and cognitive control (Derrfuss et al., 2004; Goghari & MacDonald III, 2009; Miller, 2000; Petrides, 2000). Subregions within the inferior frontal region have unique and overlapping patterns of connections with the temporal and parietal cortices, and rather than acting as a nonspecific or domain-general cognitive control mechanism, more posterior inferior frontal regions (posterior IFG and precentral gyrus) are likely more involved in non-semantic phonological control than in control of semantic processes (Gold & Buckner, 2002). The occipitotemporal region overlaps with the “visual word-form area”, which in the healthy brain appears to mediate mapping between orthographic and phonological codes (Dehaene et al., 2015; Madec et al., 2016; Mano et al., 2012). Activation of these regions, however, was not correlated with successful phonological access in our participants. This activation was numerically similar in the left hemisphere during correct and incorrect oral reading responses, suggesting that engagement of phonological rehearsal and orthographic analysis processes is only partially successful in overcoming damage to the posterior perisylvian phonological system. Activation in the right hemisphere phonological network was numerically greater for correct than incorrect trials, although this difference did not reach statistical significance.

Here we show that activation in the semantic network improves phonological retrieval for word reading in a sample of participants with phonological system damage. Given this finding, enhancement of semantic processing may be an attractive target for therapeutic interventions (Marcotte et al., 2013). Language therapies targeted at enhancing semantic retrieval are already a mainstay in patients with object naming deficits (Boyle, 2001; Efstratiadou et al., 2018). Our data suggest that such therapies may also prove beneficial for rehabilitation of oral word reading. In addition, the results indicate plausible anatomical targets for future noninvasive electrical stimulation studies in patients with phonological retrieval impairment and relatively spared semantic processing, including the left angular gyrus and left dorsomedial and inferior frontal cortex, regions that have shown previous faciliatory effects after stimulation (Campana et al., 2015; Coslett, 2016; Hamilton et al., 2011; Hartwigsen & Saur, 2019).

5. Summary and limitations

In summary, we show that neural activation in the left hemisphere semantic network, as measured with fMRI, is higher when people with phonological network damage are able to read single words aloud successfully compared to when performance on this task is unsuccessful. Although we interpret this correlation as evidence that semantic activation aids in phonological retrieval, the opposite alternative is that successful phonological retrieval aids in semantic retrieval, i.e., that semantic system activation depends on phonological retrieval. We believe this account is unlikely, however, for several reasons. First, there is ample experimental evidence from semantic priming studies that visually presented words produce rapid and automatic activation of semantic information (Carr et al., 1982; McRae & Boisvert, 1998; Neely, 2012), making it unlikely that semantic retrieval depends on phonological retrieval. Second, there is evidence that left posterior perisylvian lesions that disrupt phonology, like those in our participant sample, do not significantly impair comprehension of written or spoken words (Mesulam et al., 2015; Pillay et al., 2014), suggesting that semantic access is generally intact in such patients. Third, we did not observe differential activation between Correct and Incorrect word reading trials in the phonological network, as might be predicted if word reading success was driven by successful phonological retrieval. The correlational nature of our study, however, limits the ability to definitively exclude this alternative account, or to exclude the possibility that oral reading success depends on neural interactions between semantic and phonological retrieval processes.

As with most task-based fMRI studies in PWA, we observed wide variability in performance on the tasks, including several instances of chance performance on the phonological localizer task and one instance of chance performance on the semantic localizer task. This variability is a likely source of noise at the group level. We aimed to minimize this variability by including only correct trials in the localizer analyses, however this approach may have compromised statistical power, adversely affecting the sensitivity of these analyses. A final caveat relates to the inherent heterogeneity of PWA samples in terms of lesion size and location, which inevitably produces increased between-subject variance in behavioral and imaging analyses. We tried to minimize these effects by incorporating process-specific behavioral inclusion criteria to reduce variance in the qualitative aphasia profile, by individually tailoring the ROIs to some degree (while preserving the ability to combine data at the group level), and by incorporating extent of phonological and semantic system damage as covariates in the main analysis.