Language dominance determined by whole brain functional MRI in patients with brain lesions

Development phase (see table 1).

Right-handed volunteers. Eleven strongly right-handed, healthy volunteers (8 men, 3 women), age 19 to 43 years (mean 30), with no first-degree familial sinistrality were used because these individuals should be left dominant in nearly all cases according to published data from normal subjects19 and clinical populations.20,21

Validation phase.

Twenty-five consecutive patients (9 females and 16 males), age 16 to 70 years (mean 37) (table 2), were undergoing comprehensive evaluation at Massachusetts General Hospital (MGH) for surgical treatment of medically refractory epilepsy (2 patients) or for a brain lesion that might be close to language cortex (23 patients). Brain lesions included neoplasms (17), vascular malformations (3), heterotopia (1), sarcoidosis (1), and a benign cyst (1). All patients had preoperative EEG and routine brain MRI. Fourteen patients had Wada testing, 19 had ECS, and 8 had both Wada testing and ECS. Aphasia was noted in the history or neurologic examination as a symptom of the brain lesion or as a concomitant of focal seizures in seven patients. All subjects gave written informed consent to participate in the study, which was approved by the Human Subjects Committee at MGH. ⇓

Wada testing.

Baseline language and memory testing was administered, whenever possible, as much as 1 week before Wada testing. Cerebral angiography was performed to assess patency of the posterior communicating artery and perfusion of the posterior cerebral artery as well as to exclude patients with significant hemispheric crossflow or anomalous arterial anatomy. The hemisphere to be operated on was tested first. With arms outstretched and fingers moving, the patient counted by 2’s while 100 mg of sodium amobarbital was injected over 5 seconds into the internal carotid artery ipsilateral to the hemisphere targeted for surgery. Simultaneous multichannel EEG recorded the onset of slowing and, combined with hemiplegia of the contralateral upper limb, indicated effective anesthesia of the hemisphere. If no slowing or hemiplegia ensued, an additional 25 mg of sodium amobarbital was injected. At the onset of hemiplegia language and memory testing began and continued until return of motor function. Language testing included naming of objects from photographs and Snodgrass drawings,23 comprehension of words and sentences, and reading words and sentences. Memory testing also used naming and reading as encoding stages. After recovery from the anesthetic (30 to 45 minutes) the process was repeated on the opposite hemisphere. Language laterality was given as left, right, or equivocal dominance based on presence or absence of anomia, paraphasias, speech arrest, and errors in reading and comprehension. Determination of laterality was essentially qualitative and did not yield a numeric laterality score. Clear interruption of speech with injection of a carotid with no interruption of language with the contralateral injection despite effective anesthesia (transient hemiplegia) was the criterion used for left or right language dominance. “Equivocal” was the determination if there was no interruption of speech with either injection despite effective motor arrest. Mixed dominance was the determination if both injections resulted in impairment of language, however asymmetric the impairment.

Electrical cortical stimulation mapping.

The equipment used and methods are described in detail in a previous article.18 Initially, if exposed during craniotomy, sensorimotor cortex of the tongue and mouth were localized before language mapping to avoid confusing motor control areas with areas essential for language. With 1-mA increments in current, language areas were searched for throughout the entire exposed cortex with current ceiling determined by afterdischarges on simultaneous multichannel electrocorticography. Search for language areas involved having patients recite aloud overlearned verbal sequences—e.g., counting 1 to 20, days of the week, months of the year—with stimulation applied only after a verbal “rhythm” was established as a baseline. Visual object naming and reading words aloud were used to confirm language interruption after recitation suggested impairment. Essential language cortex was indicated by paraphasia, hesitation, or speech arrest that was reproducible with stimulation. Thresholds for language were generally higher than those for motor and sensory areas. For the 17 right-handed patients who had ECS, language impairment resulting from left-hemisphere stimulation was considered strong evidence for left dominance. Occurrence of exclusive right language dominance in right-handers is rare (approximately 4%20,24), and left-hemisphere language (revealed by ECS) is obtained in approximately 40% of right-dominant right-handers.25 Therefore, less than 2% of right-handers with left-hemisphere language foci revealed by ECS would be right dominant for language, and approximately 98% would be left dominant.

MR image acquisition.

All images were obtained using a high-speed whole body scanner (1.5-Tesla General Electric [Milwaukee, WI] Signa retrofitted for echo-planar imaging [EPI] by Advanced NMR Systems [Wilmington, MA]) and a standard RF quadrature receiver head coil. Head movement was minimized within the head coil with foam rubber pads and a Velcro strap across the forehead. Each scanning session included the following: 1) high-resolution volumetric two-dimensional sagittal spin-echo images, 2.8-mm thickness; 2) an automated shim procedure to improve magnetic field homogeneity26; 3) flow-compensated gradient-echo images colocalized with functional images; 4) high-resolution T1-weighted EPI anatomic images in-plane with functional images; and 5) blood oxygenation level-dependent (BOLD)14 contrast functional images. BOLD functional imaging used an asymmetric spin-echo sequence for T2*-weighted signal change (echo time = 70 msec, repetition time = 2 sec, offset = −25 msec, collection matrix 128 × 64, field of view = 24 cm). The asymmetric spin-echo sequence has been demonstrated to be less sensitive than standard gradient-echo sequences to large-vessel signal change.27 The resolution of functional images was 3.1 × 3.1 × 7 mm using 13 to 20 noncontiguous oblique axial slices. Throughout the duration of an experiment that lasted 2.5 or 4 minutes, 78 to 120 consecutive images per slice were acquired. Beginning with Patient 18, all patient studies were acquired using 17 slices, a 1-mm interspace, and 120 images per slice over a 4-minute experiment duration. This shift to longer experiments was done in an effort to increase temporal sampling of language cortex, which could be achieved with the advent of improved motion correction algorithms. The shorter protocol, used initially, used an oversampling mode but reduced the total available images per scan by one-half (1,024). Signal-to-noise and contrast-to-noise (functional maps) estimation between the short and long protocols did not show one to be superior over the other. The longer protocol allowed for 2,048 images per scan and the opportunity for better spatial sampling, i.e., more slices, and thus became the standard.

Measurement of lesion size.

T1-weighted high-resolution EPI images were used to manually measure the maximum dimensions of the lesion in the anteroposterior (y_l) and mediolateral (x_l) dimensions. The product of y_l and x_l was divided by the product of the maximum hemisphere extents in the anteroposterior (Y) and mediolateral (X) dimensions. A relative size index (S) was expressed as this quotient × 100.

fMRI tasks.

Stimuli were generated from a Macintosh computer in the control room with Psychlab software (Teren Gum, Montreal Neurological Institute) and projected into the magnet using an optical method. Volunteers completed three different visual language tasks—object naming, single word reading, and verb generation. Patients completed one or two repetitions of the verb generation task (no repeated items) along with other functional imaging protocols as part of their particular presurgical evaluation.

All experiments used an alternating block format with all blocks 30 seconds in duration and each trial lasting 1.5 seconds (i.e., 20 trials per block) to ensure equivalent sampling across control and activation conditions. The control task used for all experiments was fixation on a cross hair. This condition was used to minimize eye movements that can cause image artifact and to control for some bilateral V1 activation present in the activation conditions. Stimuli for the object naming task were line drawings from the Boston Naming Test,28 which were presented for a full second to allow for visual scanning of the picture. Stimuli for word reading and verb generation experiments consisted of single concrete nouns, three to seven lowercase letters in length. Words were randomly presented for 150 msec using the smallest font size easily readable by the subject to minimize activation of primary visual cortex. The remaining 1,350 msec per trial consisted of viewing the blank illuminated screen.

Participants were instructed to covertly name a pictured object (object naming), read a word (word reading), or generate a semantically related verb to a concrete noun (verb generation). An example of a trial would be the response “eat” for the stimulus “cookie.” They performed two practice trials of the verb generation task overtly before entering the scanner to ensure that they understood the task. They were told that they would be queried between experiments to estimate their performance on a given task to increase attention to the task. A covert response was used rather than an overt one because using overt responses would add undesirable image artifacts or would engender lateralized activation in the functional images. Specifically, overt speech causes both poorly correctable movement and susceptibility artifacts, whereas a manual response (e.g., keypress) could theoretically add (lateralized) processing demands in a nonequivalent manner to experimental and baseline conditions, which would be difficult to partial out post hoc (see reference 7).

fMRI data analysis.

Statistical maps comparing activation task with control task were generated voxel-by-voxel from the time series using a signed Kolmogorov-Smirnov (KS) test. Nonparametric tests such as KS may be more appropriate than parametric tests of significance such as the Student’s t-test because of aliasing of non–task-related physiologic causes of signal change conferring non-normality to the signal magnitude distribution.27 Linear drift in the baseline signal over an experiment was corrected before generation of statistical maps. A Hanning filter, which reduced the effective in-plane resolution by one-half, was applied to functional images to emphasize spatially coherent patterns of activity. Statistical maps (see figure 1) were expressed as −ln p maps using the equation where Ne is the effective number of time points.29 Images were inspected for motion artifact, which appears as a rim of increased and decreased signal along areas of high signal gradient (brain borders, ventricle margins), and images were excluded from the study if software registration methods30 failed to remove it. The time series for repeated runs of a task were combined, i.e., appended, and statistical maps regenerated on the combined time series of data. If this improved overall functional contrast-to-noise, the concatenated data set replaced the individual time series data.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1. (A) Bar graph for all right-handed volunteers on naming (light gray), reading (white), and generating tasks (dark gray). The hemisphere laterality index was greatest for the verb generation task for each subject except Subject 4, who was only tested on the naming task. (B) Bar graph showing distribution of laterality scores by task for the right-handed volunteers. Median and interquartile range are represented by the bars. Ranges are depicted by small horizontal lines above and below the bars. Note that the median and entire distribution of laterality scores for verb generation are above 19.

Quantifying laterality by fMRI.

Assessment of fMRI laterality takes into account the magnitude of signal change (as indicated by −ln p) of all voxels of a hemisphere correlating positively (−ln p > 0) with the task paradigm.31 This method resulted from the finding that, for apparently strongly left lateralized statistical maps, the left hemisphere consistently had more voxels in a given −ln p bin (where a bin is 1/100th of the full range of obtained −ln p values) than the right hemisphere above a noise threshold −ln p, which varied with each experiment (see figure 2).

• Download figure
• Open in new tab
• Download powerpoint

Figure 2. Histogram of positive −ln p values (activity positively correlated with task) for left and right hemispheres for volunteer 5 on the verb generation task demonstrating that, for a given range of −ln p) (Kolmogorov-Smirnov statistic) above a noise threshold (−ln p = 3), the left hemisphere (broken line) has more voxels than the right hemisphere (solid line). Data points were not plotted for ranges of −ln p with 0 voxels.

An activity index Ah was calculated for each hemisphere as where −ln p_i is the mean −ln p for the ith bin and n_i is the number of voxels within the ith bin.

One common laterality index has been used to assess laterality in fMRI studies of language.1,3,4 Our hemisphere laterality index (HLI) is calculated similarly using where h_l is the left hemisphere and h_r is the right hemisphere, and the denominator is the average hemisphere activity index. Thus, the HLI can theoretically range between −200 for only right-hemisphere activity and +200 for only left-hemisphere activity. Hemisphere regions of interest (ROIs) for assessing Ah were drawn on the multi-slice composite anatomic image and included the entire hemisphere and excluded the cerebellum. The threshold of −ln p > 0 was empirically chosen because using a higher threshold was shown not to systematically affect the HLI and because it allowed for a completely automated computation of the HLI after selection of the hemisphere ROI, which takes on average approximately 2 minutes per hemisphere. All computations used dedicated software from the MGH-NMR Center. For group analyses, the HLI was calculated from the statistical map generated from the concatenated data (see previous paragraph), or the algebraic average of the individual runs was used.

Development phase.

Right-handed volunteers (table 1). The HLI was left lateralized for all 10 of 10 cases for the verb generation task (see figure 1A), with a mean HLI of 32.4 (standard deviation [SD] 10) and range of 19.4 to 50.5. The reading task was left lateralized in 7 of 10 cases, with a mean HLI of 7.71 (SD 13). The naming task was left lateralized in five of seven cases, with a mean HLI of 3.8 (SD 10.5) (see figure 1B). As revealed by one-sample t-tests, the verb generation (p < 0.0001) and reading tasks (p < 0.002) were found to be significantly different from isolaterality, whereas the naming task was not (p < 0.99). On pairwise t-tests (table 3), verb generation was statistically more lateralized than either of the other two tasks [t(18), t(14), p < 0.0001], whereas reading was more lateralized than naming [t(14), p < 0.03]. In five of the six cases with data on all three tasks, the strength of laterality was verb generation > reading > naming. The occurrence of this by chance is 1/7,776 or p = 0.00013.

Left-handed volunteers (see table 1).

The HLI was left lateralized for five of eight cases for the verb generation task (figure 3A), with a mean HLI of 11.56 (SD 42) and a range of −48.77 to −62.5. The reading task was left lateralized in four of six cases, with a mean of −2.84. For all left-handed volunteers, the absolute value of the HLI was greatest for verb generation compared with reading or naming. Thus, the sign of the HLI on the verb generation task was used to divide the left-handers into a “left-dominant” (n = 5) and “right-dominant” (n = 3) group. For the left-dominant left-handers (volunteers 12, 14, 15, 17, and 19), the mean HLI for verb generation was 39.83 (SD 16), for reading 2.70 (SD 9), and object naming −0.22 (SD 9) (figure 3C). Two of four cases with data on all three tasks showed the verb generation > reading > naming order of laterality (p < 0.02). For the right-dominant left-handers (volunteers 13, 16, and 18), the mean HLI for verb generation was similar in magnitude but opposite in sign to the left-dominant groups at −35.55 (SD 20) (n = 3); reading was −30.56 (n = 1), and object naming was −5.02 (SD 9) (n = 2) (figure 3D).

• Download figure
• Open in new tab
• Download powerpoint

Figure 3. (A) Bar graph for all left-handed volunteers on naming (light gray), reading (white), and generating tasks (dark gray). For each subject, absolute value of the hemisphere laterality index (HLI) was greatest for the verb generation task compared with reading and naming tasks. (B) Bar graph showing distribution of laterality scores by task for all left-handed volunteers. Median and interquartile range are represented by the bars. Range is depicted by small horizontal lines above and below the bars. Note the wide range of laterality scores for verb generation compared with reading and naming tasks. (C) Bar graph showing distribution of laterality scores by task for the five left-handed volunteers with positive (leftward) verb generation HLI. Range is depicted by small horizontal lines above and below the bars. The small circle is an outlier. Note the similarity between this group and the group of right-handed volunteers in figure 1B. (D) Bar graph showing distribution of laterality scores by task for the three left-handed volunteers with negative (rightward) verb generation HLI. Note the comparable magnitude but opposite polarity of the distribution of verb generation scores compared with panel C.

Test-retest.

There were 15 replications of verb generation. Of the 15, three pairs were discordant (Patients 13, 15, and 23). Patient 13 had HLIs of −5.57 and 11.6; Patient 15 had HLIs of −7.79 and 5.35; and Patient 23 had HLIs of 26.24 and −2.19.

• Download figure
• Open in new tab
• Download powerpoint

Figure 4. Kolmogorov-Smirnov −ln p maps of two right-handed patients (nos. 15 and 17) performing the verb generation task. White arrows point to the stronger of homologous foci that are distributed in both hemispheres. Only slices 5 through 8 are displayed, left to right, inferior to superior. Compare the regionally mixed dominance pattern for patients 15 and 17 with studies depicted in figure 5, in which consistent left lateralized activity is the rule. Voxels that exceed the statistical threshold of −ln p = 3 (p < 0.05) are colorized according to the color scale to the right.

Comparison of patients with volunteers.

Comparing the right-handed patients with the right-handed volunteers, we used only nonaphasic right-handed patients because aphasia denotes interference with language function and potential alteration of laterality measures. The mean HLI for the right-handed patients without aphasia (mean age 33.1 years, n = 17) was 23.2 (SD 15.7) compared with 32.4 (SD 10) for the right-handed volunteers (mean age 31 years, n = 10 for the verb generation task). Statistically, the patients were slightly less lateralized than were the volunteers [p = 0.078, t(25)] and showed a trend toward greater variance [p = 0.17, F(25)]. The volunteers and patients did not differ significantly by age [p = 0.27, t(25)], nor was there an effect of age on laterality for the patients. Splitting the group of patients into a younger (mean age 26 years, n = 9) and older group (mean age 41 years, n = 8) produced mean HLIs of 24.2 and 22.1, respectively, which did not differ statistically [p = 0.79, t(15)]. Figure 5 compares results of one patient with a left parietal tumor (Patient 19) with five of the healthy right-handed volunteers on the verb generation task. What is evident in these illustrative cases is the number of common left-lateralized foci of activation extending from frontal to occipital lobes.

• Download figure
• Open in new tab
• Download powerpoint

Figure 5. Mosaic of individual case results on the verb generation task. The top 5 rows are results of healthy right-handed volunteers, and the images in the bottom row are those of a 40-year-old, right-handed man (Patient 19) with a left parietal glioma. All studies were spatially normalized into the Talairach stereotaxic reference frame32 and resliced for ease of visual comparison between studies. Voxels that exceed statistical threshold of −ln p = 3 (p < 0.05) are colorized according to the color scale to the right. These six slices were used because they demonstrate the majority of the left lateralized foci derived from averaging of the 10 right-handed volunteers on the verb generation task.8 The blue arrow points to the patient’s tumor. Images are from left to right, superior to inferior. Numeric column headings indicate axial slice location in superior-inferior dimension in millimeters relative to a line connecting anterior and posterior commissures.

Hemispheric analysis of laterality in the presence of a mass lesion.

The presence of these nine lateralized regions likely explains the predictive value of the hemispheric analysis of laterality employed in this study. For cases in which a lesion overlaps one or a few of these regions, the HLI should still be accurate because there would remain several other lateralized regions unaffected by the lesion that would contribute to the HLI. Conversely, a large lesion of the dominant hemisphere is more likely to disrupt the function of several of these regions, resulting in some degree of dysphasia. The presence of dysphasia would be prima facie evidence of a hemisphere’s indispensability for maintaining normal language function.

For the group of patients included in this study, the agreement between fMRI and invasive measures of language dominance was quite good (22 of 23). The single patient who was discordant was slightly right dominant by fMRI with HLI = −7.56, but had left-hemisphere language demonstrated by ECS and a mild nonfluent aphasia caused by a left frontal tumor. This patient had one of the largest tumors in this series of patients, involving most of the dorsolateral frontal lobe and approximately one-third of the hemisphere’s volume. The superior frontal region of brain encompassed by the tumor had very little task-related signal change (positive or negative), whereas cortex contralateral to the tumor had strong positively correlated activity (figure 6). When the tumor and contralateral cortex were omitted from analysis, left dominance was obtained (HLI = +12.55). This result suggests that this large tumor in the dominant cortex may have conferred an artifactual HLI because the remainder of the cerebrum was left dominant by the HLI. One can speculate that, with this relatively low-grade lesion of the left hemisphere, function had “shifted” to the right hemisphere, as suggested by the activity in contralateral cortex. Shifting of function to the healthy hemisphere has been reported previously in recovery from hemiparetic stroke37,38 and in aphasia39 using functional neuroimaging.

• Download figure
• Open in new tab
• Download powerpoint

Figure 6. Statistical map of Patient 20 for the verb generation task. This patient had a left frontal tumor (delineated by the green outline) that encompassed approximately one-third of the left-hemisphere volume. White arrows point to areas of strongly asymmetric activation that are present bilaterally. Note the markedly reduced number of activated voxels in the tumor region compared with the contralateral region. When the HLI was calculated, including the tumor and its reflected region in the right hemisphere, the obtained right laterality (HLI = −7.56) was inconsistent with electrical cortical stimulation result and the presence of aphasia. Exclusion of these regions from analysis yielded left laterality (HLI = +12.55). From left to right, top to bottom slices are inferior to superior. Voxels that exceed statistical threshold of −ln p = 2 (p < 0.1) are colorized according to the color scale to the right.

For this series of patients there was no significant negative correlation between lesion size and the HLI. This may be a bit misleading, however, for two reasons. First, evidence is mounting that suggests that hemisphere laterality may be a continuously distributed variable.1,8 This would have the effect of adding noise to the correlation calculation. Second, most of the lesions in this series of patients were much smaller than the lesion of Patient 20 and did not cause aphasia, suggesting that none or few of the usual lateralized language regions were disturbed by a given lesion. Because the HLI is dependent upon these lateralized regions that are relatively unaffected by a small lesion, the HLI was probably affected little in these cases. It remains to be seen how the HLI would perform with larger lesions than those in this study; however, as was true in this series, the largest perisylvian lesions would be expected to produce aphasic symptoms (Patients 6, 7, 9, 11, 17, and 20), obviating the need for other determiners of laterality.

Variability of laterality.

In our series of nonrandomly selected volunteers and patients, the majority (35/42) of subjects had strong left-hemisphere dominance. The remainder included two subjects with strong right (volunteers 13 and 16) and five subjects with more symmetric language (volunteer 18, Patients 11, 13, 15, and 17). Similar to Binder et al.,1 there appeared to be a continuum of laterality scores even within the relatively narrow range of strongly left-dominant individuals. Although it can be argued that performance differences across individuals (not independently measured) could have contributed variance to the HLIs, the obtained distribution of laterality scores is similar to that obtained by Binder et al. In their report, the reality of this interindividual variability was further substantiated by observation of a significant correlation between fMRI- and Wada-determined laterality (r = 0.61, p < 0.007) for the left-dominant patients. Two different functional anatomic models can be invoked to explain the observed continuum of language representation in the hemispheres. Each individual component function of language, e.g., phonology, orthography, semantics, syntax, etc., might be located in a single hemisphere, with the number of components in each hemisphere varying across individuals. The alternative hypothesis is that each component of language function is represented in both hemispheres but possibly more strongly in one hemisphere than the other. These two models can also operate concurrently. In a previous report8 we attempted to evaluate the importance of the first mechanism by correlating the HLI with the number of right-lateralized regions. A correlation of r = (−)0.89 was obtained (p < 0.0001), suggesting that at least the first model is operative. Newer models of the hereditability of linguistic ability and language dysfunction, e.g., dyslexia, should benefit from a better understanding of the architecture of functional neuroanatomy of language across the hemispheres elucidated using fMRI, which can survey several language components in individual subjects or across groups.

The right-handed patients were, as a group, less strongly lateralized with somewhat greater variability than the right-handed volunteers. These differences could be explained by differences in the two groups apart from the presence or absence of lesions variable in nature, size, and location. Age was not a factor because the two groups did not differ significantly by age, nor was there an effect of age on laterality within the group of patients. The right-handed volunteers were selected based on strong dextrality and absence of familial sinistrality, whereas the patient group was more heterogeneous for both. Although the relationships between handedness, genetics of handedness, and hemisphere language dominance are incompletely understood, some models hold that the presence of familial sinistrality may confer an increased likelihood of “abnormal” language lateralization.40 Accordingly, the right-handed patients might be expected to have a lower rate of clear left dominance for language.