Implementation of clinically relevant and robust fMRI-based language lateralization: Choosing the laterality index calculation method.

The assessment of language lateralization has become widely used when planning neurosurgery close to language areas, due to individual specificities and potential influence of brain pathology. Functional magnetic resonance imaging (fMRI) allows non-invasive and quantitative assessment of language lateralization for presurgical planning using a laterality index (LI). However, the conventional method is limited by the dependence of the LI on the chosen activation threshold. To overcome this limitation, different threshold-independent LI calculations have been reported. The purpose of this study was to propose a simplified approach to threshold-independent LI calculation and compare it with three previously reported methods on the same cohort of subjects. Fifteen healthy subjects, who performed picture naming, verb generation, and word fluency tasks, were scanned. LI values were calculated for all subjects using four methods, and considering either the whole hemisphere or an atlas-defined language area. For each method, the subjects were ranked according to the calculated LI values, and the obtained rankings were compared. All LI calculation methods agreed in differentiating strong from weak lateralization on both hemispheric and regional scales (Spearman's correlation coefficients 0.59-1.00). In general, a more lateralized activation was found in the language area than in the whole hemisphere. The new method is well suited for application in the clinical practice as it is simple to implement, fast, and robust. The good agreement between LI calculation methods suggests that the choice of method is not key. Nevertheless, it should be consistent to allow a relative comparison of language lateralization between subjects.

Introduction

In the majority of individuals, language functions are predominantly located in the left brain hemisphere [1]. Nevertheless, language centers are usually spread over both hemispheres and vary from individual to individual in both their location and extent [2]. In addition, hemispheric or regional dominance in language functions (language lateralization) has been shown to depend on several factors such as age, handedness [3,4] and varies when evaluated in different brain regions [5,6]. More importantly, brain pathology can alter intra- and interhemispheric brain functionality [7], potentially causing a reorganization of language centers. Such a reorganization has been shown to lead to a change of language laterality from left to right hemisphere in left temporal lobe epilepsy [8-11], explaining the higher occurrence of atypical language lateralization (i.e. not left dominant) in epilepsy patients [12]. A similar change of language laterality has also been demonstrated in patients recovering from a stroke [13], and in brain tumor patients [4]. For these reasons, the assessment of language lateralization has become widely used when planning neurosurgery close to language areas. The information obtained from such assessment is useful for estimating the risk of post-surgical deficits [14], selecting an appropriate surgical approach [15,16], and deciding on the resection extent [16,17,18]. Direct electrical stimulation and the intracarotid (sodium) amobarbital test or Wada test [19] have often been used to assess language lateralization. However, these methods are invasive and can lead to complications and irreversible disabilities [20,21].

Functional magnetic resonance imaging (fMRI) can also be used for the assessment of language functions and has the advantage of being non-invasive, quantitative and suitable for presurgical planning. The assessment of language lateralization using fMRI involves the comparison of signals obtained during rest and activation phases of a task performed by the subject during the scan. This signal comparison is performed at the voxel level using a statistical test and produces activation maps, known as statistical parametric maps (SPMs). Hemispheric or regional dominance in language functions can then be quantified employing the laterality index (LI), which indicates the prevalence of activation in one side of the brain over the other [17,22]. The LI is conventionally calculated using the following formula: where NL and NR are the number of voxels with value above a specific activation threshold in left and right region of interest (ROI), respectively. LI values thus range from +1 (left dominant) to -1 (right dominant). However, a major limitation of this approach is the strong dependence of the LI on the arbitrarily chosen activation threshold value [23-25].

To overcome this limitation, various threshold-independent laterality index calculation methods have been reported in the literature [25]. For example, Knecht and colleagues define the activation threshold by a fixed total number of voxels and from this calculate a single LI value [26]. Abbott and colleagues expanded this idea and proposed to calculate the LI as a function of the total number of activated voxels and produce a LI curve, which can be used to visually evaluate a patient’s lateralization compared to a healthy control group [24]. Matsuo and colleagues proposed to calculate a global LI by averaging conventional LI values evaluated over a range of thresholds [27]. Branco, Suarez and colleagues proposed to integrate the weighted histogram of voxel counts against threshold in order to calculate a global LI [23,28]. The review by Bradshaw and colleagues [25] highlights the lack of standardization in assessment of language lateralization using fMRI. The multitude of LI calculation methods, tasks and ROIs used in previously published reports makes comparison of results between studies difficult. Furthermore, the implementation of these methods is not always straightforward, which limits the feasibility of adoption in the clinical routine. When wishing to implement the assessment of language lateralization using fMRI in the clinical routine, a choice has to be made regarding the method. A direct comparison between threshold-independent LI calculation methods using the same subjects, tasks, and ROIs could provide useful insights to aid this choice.

In this work we 1) implement and compare three previously reported threshold-independent methods applied to the same set of fMRI datasets from healthy volunteers, and 2) propose a simplified approach to threshold-independent LI calculation (AUCLI), which we compare to the other methods. Lateralization is evaluated on both a hemispheric and regional scale using three different language tasks.

Methods

MRI sequence protocol

Fifteen healthy right-handed volunteers (age range 21–45, mean ± standard deviation = 29 ± 7, 12 female) were scanned at 1.5 T (Magnetom Aera, Siemens AG, Erlangen, Germany) using the standard 20-channel head-only receive coil. Informed written consent was obtained with ethical approval from the UK National Research Ethics Service (REC 04/Q0706/72). The MRI sequence protocol consisted of a 3D T1-weighted MPRAGE anatomical sequence (TE/TR = 3.02/2200 ms, voxel = (1 mm)3, FA = 8°, parallel imaging acceleration of 2) and three BOLD contrast fMRI gradient echo EPI sequences (TE/TR = 40/3000 ms, voxel = 2.5x2.5x3 mm3). For a given language paradigm, the fMRI protocol consisted of 6 cycles of alternating rest and activation periods of 30 seconds, resulting in a total scan time of 6 minutes (120 measurements).

fMRI paradigms

The stimuli consisted of black letters or drawings on a white background and were presented visually to the subject using a screen at the end of the scanner bed, visible via a set of mirrors positioned on the head coil. Each volunteer performed three different language tasks: verb generation, picture naming, and word fluency. For verb generation, nouns appeared on the screen (15 per cycle) and the subject had to silently generate verbs associated with the noun. For the picture naming task, line drawings appeared on the screen (10 per cycle) and the subject had to silently name the depicted object. For the word fluency task, letters appeared on the screen (7 per cycle) and the subject had to generate words starting with the presented letter. For each task, the stimuli were randomly chosen from a pool of nouns, pictures, or letters. During the resting periods, a black cross-hair on a white background was projected in the center of the screen. The language tasks were set up using SuperLab 4.0 (Cedrus, San Pedro, California, USA).

Image processing

The fMRI data was processed with the software package SPM12 (Wellcome Trust Centre for Neuroimaging, University College London, UK) using an in-house developed batch processing pipeline, which included following steps: 1) Images within each fMRI dataset were realigned to the first image (rigid body spatial transformation and least square algorithm, SPM12) to compensate for small degrees of motion. 2) The fMRI data was then co-registered to the anatomical data (non-linear mutual information registration algorithm, SPM12) and 3) smoothed using an isotropic Gaussian kernel with 8 mm full width at half maximum. SPMs were then calculated within the general linear model framework using a Student’s t-test with a family-wise error rate significance level set at 0.05. In the following, the voxel values of the calculated SPM are referred to as threshold values or t-values. For all LI calculations, only voxels with positive t-values are considered as these indicate activation correlating with the performed task.

Regions of interest

In this work, the LI was calculated using both hemispheric and regional ROIs. For the definition of the ROIs, different brain atlases available in FSL (Wellcome Centre for Integrative Neuroimaging, FMRIB Analysis Group, University of Oxford, UK) were employed. The high resolution T1-weighted Montreal Neurological Institute (MNI) standard brain was used as reference and was mapped to the T1-weighted MPRAGE acquisition using both affine and non-linear registrations (FSL FLIRT [29-31] and FNIRT [32,33], respectively), thus allowing the direct transformation of atlas structures from the standard space to the acquisition space. Left and right hemisphere ROIs encompassing the entire cortical hemispheres (excluding the cerebellum) were created from the anatomical structures defined in the Harvard-Oxford Subcortical Structural Atlas [34-37]. The language ROIs encompassed Broca’s area (Brodmann’s areas 44 and 45) and Wernicke’s area (posterior division of the superior temporal gyrus) as defined by the Jülich Histological Atlas [38-41] and Harvard-Oxford Cortical Structural Atlas [34-37]. The ROIs defined in the acquisition space for one of the subjects can be seen in Fig 1.

Fig 1

Regions of interest defined using standard brain atlases.

1 and 2 are the right and left hemisphere ROIs, 3 and 4 are the left and right language ROIs with ‘a’ designating Broca’s area and ‘b’ Wernicke’s area.

Threshold-independent laterality index calculation methods

All LI calculations were performed using in-house software developed in MATLAB (Version 2016a, The MathWorks, Inc., Natick, Massachusetts, USA).

1. Fixed total number of activated voxels (curveLI)

The method reported by Abbott and colleagues [23], labeled as curveLI, relies on the fact that each threshold corresponds to a total number of activated voxels. The curveLI method plots the LI values calculated for each threshold using Eq (1) against the total number of activated voxels. The total number of activated voxels considered for the calculations ranged from zero (no voxel within the ROI) to the maximum (all voxels with positive value within the ROI). In order to plot the curves obtained for different subjects in a single graph, the total number of activated voxels was normalized to the maximum for each subject, bringing the x axis in the range 0 to 1. This compensated for differences in ROI size from subject to subject due to individual brain anatomies. The distributions from all subjects were then used to calculate a mean LI and the 95% confidence interval (CI) of this mean:

Where curveLIindividual denotes the LI calculated for a single individual, and n denotes the total number of subjects (n = 15 in this study). A single LI value was extracted at a threshold value corresponding to half the voxels being active (mid curve) for comparison with the other LI indices.

2. Average (AveLI)

Following the method reported by Matsuo and colleagues [26], LI values were calculated for each t-value existing within the ROI considered, using Eq (1). These individual LI values were then averaged to form a global index:

Where tmin and tmax are the minimum and maximum t-values, and N is the sum of left and right total number of voxels with positive t-values.

Weighted histogram (histoLI)

Following the method reported by Branco, Suarez and colleagues [22,27], histograms of voxel counts versus t-value were determined for left and right ROIs. The t-values used for the histograms ranged from 0 to the maximal t-value within the ROI considered and were binned using an automatic binning algorithm, yielding bins with uniform width covering the whole range of t-values (MATLAB histcounts). A weighting function of squared t-values was then used to weight these voxel histograms [27]. The areas under the curves of left and right weighted histograms were calculated using a trapezoidal numerical integration, and compared as follows: where LA and RA are the areas under the weighted histograms in left and right ROI, respectively.

Area under the curve (AUCLI)

Another way of looking at the difference in activation in left and right ROI is to consider the number of voxels with values above the threshold for all possible thresholds [42]. This is achieved by setting every t-value present in the ROIs as a threshold and record the number of voxels with value above the threshold. The cumulative histograms of the obtained number of voxel vs threshold for left and right ROI (shown in Fig 2), can then be compared. The new method we propose in this work calculates the areas under these histograms and quantitatively compares them using the following formula: where AUCL and AUCR are the areas under the cumulative histogram calculated for the left and right ROI, respectively. The areas under the cumulative histograms were calculated using the trapezoidal numerical integration available as a function in MATLAB.

Fig 2

Calculation of the laterality index with the AUCLI method.

The area under the cumulative histograms of number of voxels with value above threshold for left and right ROIs are used to compute a laterality index for the novel AUCLI method.

Method comparison

LI values were calculated for each fMRI dataset using the four different threshold-independent LI calculation methods. For each method, the subjects were ranked according to the calculated LI values, and pair-wise comparisons of rankings were performed for each task and ROI combination. To quantify the agreement between pairs of rankings, Spearman’s correlation coefficients ρ were calculated using SPSS (Version 24, Statistics Software, IBM Corporation, USA).

Results

All calculated LI values can be found as supplementary data in S1 Table.

LI versus threshold plots resulting from the conventional LI calculation are shown in Fig 3 for the three tasks and both ROIs. These plots demonstrate that the LI varies over the range of threshold, highlighting the existing issue of the conventional LI calculation, namely having to choose a single threshold value to evaluate lateralization. The LI versus threshold curves are not stable over the range of thresholds and can have sudden variations (drops), showing that the degree of lateralization can change considerably with the threshold. The 95% confidence interval of the mean LI is larger for the picture naming task than for the two other tasks, indicating that the differences between subjects are greater for this task. Furthermore, as both task and ROI have an impact on the LI distributions, the same subject can have different degrees of lateralization in different tasks or ROIs.

Fig 3

Conventional laterality index plots.

The dependence of the laterality index on the threshold varies with the subject, the task performed, and the ROI chosen.

The dependence of the LI on the total number of activated voxels (curveLI method) also varies with subject, task, and ROI (see Fig 4). Nevertheless, the LI has smoother variations over the range of total number of activated voxels than over the range of thresholds, as expected [24]. Fig 4 shows variability between tasks similar to Fig 3: the spread in individual subject curves is greater for the picture naming task than for the two other tasks when considering the language ROI, and the lateralization of some subjects differs between tasks. Overall, the curves obtained with the curveLI method show higher LI values for the language ROI than for the hemisphere ROI (see Fig 4). This indicates a more lateralized activation on a regional scale (language ROI) than on a larger scale (hemisphere ROI). However, within each task, the highest and lowest curves belong to the same subjects for both ROIs.

Fig 4

Laterality index versus total number of activated voxels (curveLI) plots.

The dependence of the laterality index on the total number of activated voxels varies with the subject, the task performed, and the ROI. The data points for the mean LI and 95% confidence interval can be found as supplementary data in S2 Table.

Fig 5 compares the inter-subject distributions of LI values in each method, for each task and ROI. The boxplots show higher LI values for the language ROI than for the hemisphere ROI in all methods. The median LI values obtained for the histoLI method are visibly discrepant from the other three methods. Fig 6 shows a different aspect of the same data, visually comparing the subject rankings obtained with each method. Within each graph, a color gradient (from blue for low LI values to red for high LI values) was applied, based on the subject ranking from the curveLI method (left side of the plot). In this visualization, crossing lines indicate differences in ranking between the methods. Fig 6 shows that, overall, similar subject rankings were obtained across the four methods, with Spearman’s correlation coefficients ranging from 0.59 to 1.00 (Table 1). The pairwise comparisons involving the histoLI method resulted in the lowest Spearman’s correlation coefficient. For the hemisphere ROI, the agreement between the AveLI and AUCLI methods is optimal in all tasks, while for the language ROI, the curveLI and AUCLI methods showed the best agreement in all tasks. Different tasks led to different subject rankings, but the level of agreement (i.e. range of Spearman’s correlation coefficients) between methods in terms of subject ranking is similar between tasks. Within each task, there are some variations in the ranking between hemispheric and language ROIs, but weak or strong laterality is preserved for most subjects. The agreement between subject rankings obtained with the same method for different ROIs is good, with Spearman’s correlation coefficients ranging from 0.75 to 0.96 (Table 2).

Fig 5

Boxplots showing the variation between subjects within each method.

For each box, the central line indicates the median. The bottom and top edges indicate the first quartile q1 and third quartile q3, respectively. The whiskers extend to the most extreme data points not considered as outliers. Outliers are shown individually as ‘+’ on the plots. Outliers are hereby defined as values larger than q3+w(q3-q1) or smaller than q1-w(q3-q1), where w is the maximum whisker length.

Fig 6

Visual comparison of the subject rankings obtained with the different methods.

Within each graph, a gradual color change from blue for low LI values to red for high LI values is attributed to each subject, reflecting the subject ranking obtained with the curveLI method (left side of the plot). Crossing lines indicate differences in subject rankings between the methods.

Table 1

Spearman’s correlation coefficients for comparisons of subject rankings between the four different LI calculation methods for both ROIs.

	picture naming task					verb generation task					word fluency task
language ROI		curveLI	AveLI	histoLI	AUCLI		curveLI	AveLI	histoLI	AUCLI		curveLI	AveLI	histoLI	AUCLI
	curveLI	1.00	0.96	0.9	0.99	curveLI	1.00	0.99	0.68	1.00	curveLI	1.00	0.96	0.79	0.97
	AveLI	0.96	1.00	0.98	0.97	AveLI	0.99	1.00	0.65	0.99	AveLI	0.96	1.00	0.85	0.99
	histoLI	0.90	0.98	1.00	0.92	histoLI	0.68	0.65	1.00	0.71	histoLI	0.79	0.85	1.00	0.85
	AUCLI	0.96	0.97	0.92	1.00	AUCLI	1.00	0.99	0.71	1.00	AUCLI	0.97	0.99	0.85	1.00
hemisphere ROI		curveLI	AveLI	histoLI	AUCLI		curveLI	AveLI	histoLI	AUCLI		curveLI	AveLI	histoLI	AUCLI
	curveLI	1.00	0.94	0.80	0.94	curveLI	1.00	0 98	0.59	0.96	curveLI	1.00	0.96	0.78	0.97
	AveLI	0.94	1.00	0.93	0.96	AveLI	0.98	1.00	0.63	0.99	AveLI	0 .96	1.00	0.83	0. 99
	histoLI	0.80	0.93	1.00	0.86	histoLI	0.59	0.63	1.00	0.67	histoLI	0.78	0.83	1.00	0.84
	AUCLI	0.94	0.96	0.86	1.00	AUCLI	0.96	0.99	0.67	1.00	AUCLI	0.97	0.99	0.84	1.00

Similar subject rankings were obtained across the four methods for both ROIs, with Spearman’s correlation coefficients ranging from 0.59 to 1.00.

Table 2

Spearman’s correlation coefficients for comparisons of subject rankings between the language ROI and the hemisphere ROI for the four different LI calculation methods.

picture naming task				verb generation task				word fluency task
curveLI	AveLI	histoLI	AUCLI	curveLI	AveLI	histoLI	AUCLI	curveLI	AveLI	histoLI	AUCLI
0.75	0.84	0.85	0.78	0.96	0.91	0.75	0.88	0.90	0.86	0.90	0.85

The agreement between subject rankings obtained with the same method for the language ROI and the hemisphere ROI is good, with Spearman’s correlation coefficients ranging from 0.75 to 0.96.

Discussion

In this work, we compare four different threshold independent methods for assessment of language lateralization with fMRI. To eliminate the problem of the dependence of the LI on the statistical threshold, the different approaches adopt different metrics and it is therefore important to establish if, in practice, the choice of method has an impact on the assessment of language lateralization. A direct comparison by applying the methods to the same subject cohort represents a useful evidence base for the implementation of robust fMRI-based language lateralization in the clinical routine. This work has found that overall there is a good correlation between different methods in terms of subject rankings.

curveLI method

The method based on choosing a fixed total number of activated voxels (curveLI) attempts to give an objective assessment of a patient’s language lateralization in relation to a healthy control group [24], and is useful for looking at variations of the LI both within a subject and between subjects. Our results show that the curveLI method, compared to conventional LI assessment, reduced variability within subject. This matches the results from Abbott and colleagues, despite the differences in fMRI acquisition (MRI scanner, sequence parameters, tasks) and ROI definition between their study and ours, as well as our choice of normalization for the total number of activated voxels, highlighting the robustness of this method against the above-mentioned factors. However, there are visible differences between the mean LI value and 95% confidence interval obtained in our study for the word fluency task and those reported by Abbott and colleagues [24]. One limitation of this method is therefore that the reference set of curves might be specific to the local implementation and thus not transferrable to other datasets.

AveLI method

The AveLI method produces a global laterality index value taking into account the lateralization at each activation value in the data series, giving more weight to voxels with high activation value compared to those with low activation [27]. The AveLI method has been shown to be resistant to outliers and stable against noise, to yield highly reproducible LI values between tasks, and to allow a good separation of subjects into left, right or bilateral language lateralization categories [27]. However, our results show that the LI values obtained with the AveLI method yield different subject rankings in different tasks. This might be due to differences in the type of tasks used. While we used picture naming, verb generation and word fluency tasks in English, Matsuo and colleagues used word generation and homophone judgement tasks in Chinese language.

histoLI method

The reduced agreement between the histoLI method and the other three is likely due to the choice of the weighting function. We adopted a weighting function of squared t-values as recommended by Suarez and colleagues [28] for assessing language lateralization, instead of a linear weighting function as reported in their previous work looking at presurgical assessment of memory lateralization in the hippocampus [23]. This choice was motivated by the fact that language areas encompass a larger volume than the hippocampus [28]. In larger volumes, the probability of including voxels presenting low activation is higher so that a weighting function, which further increases the impact of voxels with high activation, is more appropriate. The weighting function of squared t-values reduces the influence of low t-value voxels (noise and false positives), which should improve the accuracy of the results, but also increases the influence of high t-value voxels. This overall results in a widened range of LIs, both towards the highly positive and the negative values (see Fig 5).

AUCLI method

The novel method we propose in this work produces a single LI value evaluated over the entire range of activations present in the data series. The LI values calculated with the ACULI method are in good agreement with the other methods, especially with curveLI and AveLI. Nagata and colleagues [42] proposed another interesting LI calculation method. Their idea was to fit the same monomial equation to the number of activated voxels versus threshold curves for left and right ROI. The LI can then be calculated by comparing the fit parameters obtained for left and right ROI. When this method was applied to our data, it was not possible to find a common function that would satisfactorily describe both the left and right voxel histogram curves. Both voxel histogram curves shown in Fig 2, could be fitted with an exponential or bi-exponential function but not with the same monomial function. A comparison of fit parameters to remove the direct dependence of the LI on the activation threshold as Nagata and colleagues [42] suggested, was thus not possible with such curves. The method we propose here is a simplification of Nagata’s method: instead of comparing the two distributions using a fit function, we measure the area under the curves, which is independent of the shape of the curve. This approach also overcomes the drawback of having to find the best fit function, which may be specific to the task or ROI used.

Suitability for the clinical routine

To be suitable for the clinical routine, a LI calculation should be 1) robust (i.e. independent of any parameter), 2) reproducible (i.e. stable over multiple calculations) and 3) allow an easy subject comparison [42]. The curveLI method and the AveLI method fulfil these three criteria. The histoLI method satisfies criteria 2) only for a determined weighting function, as acknowledged by Branco and colleagues [23] who first introduced this method. The AUCLI method evaluates the LI over the entire range of activation thresholds (t-values) present in the activation map, making it independent of the activation threshold and therefore robust. The AUCLI method also satisfies criteria 2) as the calculations of the cumulative histograms and areas under these will yield identical results each time performed, and therefore the final LI calculation will be stable over repeated calculations. All presented threshold-independent methods can provide a single summarizing LI value, which makes comparison between subjects easy, therefore satisfying criterion 3). However, the curveLI method also offers visual comparison of curves plotted in a single graph (Fig 4).Such a comparison is valid only if the number of activated voxels is appropriately normalized (for example to the total number of voxels with positive values within the SPM for each subject, as done in this work), or if the total number of positive voxels within the ROI considered is the same for all subjects, as done by Abbott and colleagues [24]. When a reference cohort is available, the curveLI method offers an easy visual comparison (direct assessment of where the patient curve lies in respect to the mean LI curve and 95% confidence interval), while retaining the more complex information of the smooth dependence of the the LI on the total number of activated voxels. In addition to the three criteria listed by Nagata and colleagues [42], 4) ease of implementation and 5) speed of data analysis are important for use in the clinical routine. All four methods investigated in this study require custom scripts, but the implementation of the AUCLI method is more straightforward as the metric used is simpler than in the other methods. For a single subject, the LI calculation required approximatively 10 seconds with the curveLI method, 60 seconds with the AveLI method, 5 seconds with the histoLI method, and 1 second with the AUCLI method using our computer environment (Mac OS X El Capitan Version 10.11.6, Processor 3.3 GHz Intel Core i5, RAM 32 GB 1867 MHz DDR3). Taking these five criteria into account, the novel method proposed in this work can be considered suitable for the clinical routine and presents advantages compared to the previously reported methods.

Choice of ROI

In this work, two different ROIs have been considered, giving the possibility to assess language lateralization on both a hemispheric scale and a more regional scale. Even though the agreement between methods is slightly higher when the language ROI is considered rather than the hemisphere ROI, choosing the language ROI defined using standard brain atlases might not be appropriate in clinical subjects. Problems with the registration of the MNI standard brain to the acquired brain might arise due to abnormal brain anatomy. Furthermore, the actual functional area of the patient might not correspond to the atlas-based language ROI even if the registration yields satisfactory results. In patients with abnormal anatomy and potentially displaced functional centers, it might thus not be possible to accurately define functional ROIs based on anatomical or functional a-priori knowledge. In such cases, which are common in presurgical assessment of language lateralization, larger ROIs encompassing the whole hemisphere may be preferable at the sacrifice of including areas that are not associated with language. For instance, strong activation was visible in the visual cortex for all subjects as a result of the visual stimuli. This activation is not perfectly symmetric and can thus influence the calculated LI values. This problem could be overcome by masking out the visual cortex to exclude this area of the brain from the calculation and remove its influence on the LI values or by modifying the paradigm design to obtain similar visual activation during resting and active part of the task. Another drawback of using a hemispheric ROI to assess language lateralization is the impossibility to describe the differences in lateralization between different regions within the brain hemispheres. Regional language lateralization has been shown to differ from hemispheric language lateralization in both healthy subjects and patients [6]. The information about a regional lateralization might be of greater interest than a hemispheric lateralization in case of a very localized lesion. The choice of hemispheric or regional ROI should thus be made according to the patient’s condition.

Choice of language tasks

In this work, we have considered three language tasks. The results show that the assessment of language lateralization is task dependent, confirming previous conclusions [43-45]. This is a result of the complexity of human language, which involves numerous different functions [1,2,45]. Not all language functions can be assessed by a single fMRI task and it is therefore recommended to use an assembly of tasks to assess language lateralization more accurately [2,5,45]. The LI values calculated for the picture naming task are noticeably lower than those obtained for the verb generation and word fluency tasks (see Fig 6). Our results thus confirm that the picture naming task is less reliable than the two other tasks considered in this work. However, the picture naming task is simpler to perform and may therefore be more appropriate for patients with cognitive deficits [45].

Limitations

The multilingual character of the cohort of healthy subjects used in this project influences the range of LI values obtained since activation patterns have been shown to differ in native speakers and non-native speakers [46], potentially leading to a more spread out confidence interval of the mean LI values. However, a multilingual control group might be more representative of the expected patient population in certain hospitals.

The inherent limitation of fMRI resulting from poor patient cooperation and task performance should always be kept in mind. Especially when using silent language tasks, it is very difficult to ensure the patient is performing the task properly during the scan. Therefore, it is important to give clear instructions before the scan and some practice outside the scanner might even be advisable. For patients with cognitive deficits it may be necessary to verify whether the tasks can be performed before the scanning session. Some adaption of the task design (e.g. color of writing and background, display during rest period, frequency of image/word/letter) might also be useful for clinical cases to improve the patient’s ability to perform the task.

This work based on volunteer data, was part of the process to establish fMRI-based presurgical assessment of language lateralization in the clinical routine at our hospital, and is focused on the comparison of different LI calculation approaches. Healthy volunteer data, in addition to provide a reference cohort, is ideally suited for direct method comparison as they are independent of differences in patient pathologies. As patients are now being considered for presurgical assessment of language lateralization, future developments of this work will be to compare the methods reported here in clinical subjects, and further investigate the choice of hemispheric vs language ROI. In these subjects, comparison of the fMRI results to direct electrical stimulation during surgery–the gold standard for determining lateralization–will also be possible.

Conclusion

For robust assessment of language lateralization in the clinical routine, it is advisable to use a threshold-independent laterality index calculation. In this work, we have tested four different methods on the same subject cohort. Our results show that the choice of method itself is not key, as all methods agree in differentiating strong from weak lateralization on both hemispheric and regional scales. This choice should nevertheless be consistent to allow a relative comparison of language lateralization between subjects. Our results highlight that the laterality index is not an absolute measure, as numerous factors—some purely related to the LI calculation method—can influence its value. In this work, we have introduced a new threshold-independent laterality index calculation method and validated it against three previously reported methods. Our evaluation suggests that the new method is well suited for application to clinical practice as it is simple to implement, fast, robust, reproducible, and allows direct subject comparison.

LI values calculated for all subjects with the four threshold-independent LI calculation methods for all tasks and both ROIs.

(XLSX)

Click here for additional data file.

Mean LI value and 95% confidence interval data points obtained with the curveLI method.

(XLSX)

Click here for additional data file.

12 Nov 2019

PONE-D-19-25144

Implementation of clinically relevant and robust fMRI-based language lateralization: choosing the laterality index calculation method

PLOS ONE

Dear Dr De Vita,

Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process.

The reviewers suggested a number of clarifications and extensions of the manuscript, please respond as required.

We would appreciate receiving your revised manuscript by Dec 27 2019 11:59PM. When you are ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter.

To enhance the reproducibility of your results, we recommend that if applicable you deposit your laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. For instructions see: http://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols

Please include the following items when submitting your revised manuscript:

A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). This letter should be uploaded as separate file and labeled 'Response to Reviewers'.

A marked-up copy of your manuscript that highlights changes made to the original version. This file should be uploaded as separate file and labeled 'Revised Manuscript with Track Changes'.

An unmarked version of your revised paper without tracked changes. This file should be uploaded as separate file and labeled 'Manuscript'.

Please note while forming your response, if your article is accepted, you may have the opportunity to make the peer review history publicly available. The record will include editor decision letters (with reviews) and your responses to reviewer comments. If eligible, we will contact you to opt in or out.

We look forward to receiving your revised manuscript.

Kind regards,

Peter Lundberg

Academic Editor

PLOS ONE

Journal Requirements:

1. When submitting your revision, we need you to address these additional requirements. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at http://www.journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and http://www.journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf

2. In your data availability statement you write all the relevant data are within the paper and/or its Supporting Information files. Please ensure you have provided the individual data points used to create the figures and determine means, medians and variance measures presented in the results, tables and figures (http://journals.plos.org/plosone/s/data-availability#loc-faqs-for-data-policy). If these data cannot be publicly deposited or included in the supporting information, e.g. due to patient privacy or ownership by a third party, explain why and explain how researchers may access them.

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

Reviewer #2: Yes

**********

2. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: Yes

**********

3. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: No

Reviewer #2: No

**********

4. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: Yes

**********

5. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: Thank you for a well-written article on an interesting topic, keeping the useability of the results for clinicians and researchers in mind throughout the paper. I have some minor comments.

I cannot find the full dataset (fMRI raw data or statistical parameter maps) provided anywhere as attachment, supplementary data or link to data repository.

Line 168/169 on page 8 would read better if the numbers and letters referred to were put in quotation marks, e.g. "... with 'a' designating Broca's area"

The description of the novel method, AUCLI, was lacking some details that would aid reproducibility of the method. First, line 213-215 reads "... count the number of voxels with values above the threshold for all t-values present in the ROIs". Is this meant to apply to each threshold present in the ROIs? The sentence reads as if one threshold is set, but this cannot be the case as the manuscript is advocating the use of threshold-independent LIs.

The first sentence of the Results, at page 11 line 239 refers to LI versus threshold plots that are likely referring to the conventional LI. Please do state this more clearly in this sentence. As LI is not consistently used only to refer to the conventional LI calculation, it is helpful to be very consistent in naming. Likewise, in the description of Fig 4 it is helpful to refer to the in-manuscript name of the method; curveLI.

In the figure text of Fig 5 (page 13 lines 295-296) there's a reference to outliers. The manuscript is lacking a clear description of outlier detection chosen, and if there is need for outlier detection (rationale, implications).

At page 17/18 there is not a clear positive answer to whether other methods than CurveLI adhere to criterion 3 for suitability for clinical implementation: "allow an easy subject comparison". Therefore, the conclusion drawn in line 412/413 "Taking these five criteria into account, the novel method proposed in this work can be considered suitable for the clinical routine" cannot be drawn which undermines one of the main findings of the manuscript. Of course, the authors do show means of comparing between subjects throughout the manuscript, so this should be described in this section together with a judgment of whether this is adequate enough to be deemed as "easy", or else rephrase their conclusions regarding clinical implementability.

It should be noted that the colors and gradient used in Fig 6 are virtually indiscriminable when viewed as a black and white printout, it would be advisable to switch to colors with a clear difference in b/w.

Reviewer #2: The authors present a study comparing methods for calculating laterality index. While the findings may not have broad appeal, they will certainly be of use for groups performing clinical fMRI for pre-surgical mapping.

I have a number of comments, though.

1. Why was such a slow TR of 3 seconds chosen? It has been apparent for a while that the sensitivity of fMRI increases with decreasing TR.

2. When using SPM, it is generally recommended to choose a smoothing kernel that is approximately 2x the size of the acquired voxel (see Friston et al., 1996 NeuroImage; Ball et al., 2012 Human Brain Mapping; Pajula and Tohka 2014, MRI; Liu et al., 2017, J Neuroscience Methods). Given that the acquired voxel is 2.5x2.5x3mm, can the authors justify using an 8mm smoothing kernel, which is almost 3x the size of the acquired voxel?

3. What did the authors include in their GLM beyond the task regressors? Were the realignment parameters estimated during the realignment step included as covariates of no interest? What steps were taken to mitigate the effect of noise in the data? What quality checks were performed on the MRI data? Individual examination for excessive head motion?

4. I do not understand the purpose of Figure 6. What is it showing that could not be shown in Figure 5, if the authors included the individual data points as part of the box and whisker plots in Figure 5? The authors need to better explain what is being shown in Figure 6, especially why it is important to color code the subjects based on curveLI method?

5. Did the authors consider acquiring multiple runs of each task in a single session or performing multiple scanning sessions to enable any sort of test-retest or cross-validation of the results? This would be an interesting result to see in terms of how robust each LI method is.

6. In the conclusions, the authors mention that laterality index measures can be influenced by post-processing method. Given the results presented in the manuscript, how did the authors arrive at this conclusion? I did not see any indication that the authors varied anything but whether LI was calculated at the hemispheric level or using language-related ROIs. To justify this conclusion, I would have expected to see the authors comparing smoothing kernels, standard software packages/pipelines, the inclusion/exclusion of noise confounds in the GLM, etc.

7. The authors have omitted many of the required references and acknowledgments for using the Harvard Oxford and Julich Histological atlases (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/Atlases).

**********

6. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files to be viewed.]

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email us at figures@plos.org. Please note that Supporting Information files do not need this step.

17 Dec 2019

A full response to reviewers comments is attached as a separate file

Submitted filename: ResponseToReviewers_final_R1.docx

Click here for additional data file.

24 Feb 2020

Implementation of clinically relevant and robust fMRI-based language lateralization: choosing the laterality index calculation method

PONE-D-19-25144R1

Dear Dr. De Vita,

We are pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it complies with all outstanding technical requirements.

Within one week, you will receive an e-mail containing information on the amendments required prior to publication. When all required modifications have been addressed, you will receive a formal acceptance letter and your manuscript will proceed to our production department and be scheduled for publication.

Shortly after the formal acceptance letter is sent, an invoice for payment will follow. To ensure an efficient production and billing process, please log into Editorial Manager at https://www.editorialmanager.com/pone/, click the "Update My Information" link at the top of the page, and update your user information. If you have any billing related questions, please contact our Author Billing department directly at authorbilling@plos.org.

If your institution or institutions have a press office, please notify them about your upcoming paper to enable them to help maximize its impact. If they will be preparing press materials for this manuscript, you must inform our press team as soon as possible and no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org.

With kind regards,

Peter Lundberg

Academic Editor

PLOS ONE

Additional Editor Comments (optional):

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation.

Reviewer #1: All comments have been addressed

Reviewer #2: All comments have been addressed

**********

2. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

Reviewer #2: Yes

**********

3. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: Yes

**********

4. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: Yes

Reviewer #2: Yes

**********

5. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: Yes

**********

6. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: Thank you for your clarifications, all my concerns have been adequately addressed.

I would advice to include a statement about the described restrictions in data sharing in the article.

Also, note that the word 'histograms' is accidentally written twice in line 217

Reviewer #2: (No Response)

**********

7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

2 Mar 2020

PONE-D-19-25144R1

Implementation of clinically relevant and robust fMRI-based language lateralization: choosing the laterality index calculation method

Dear Dr. De Vita:

I am pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department.

If your institution or institutions have a press office, please notify them about your upcoming paper at this point, to enable them to help maximize its impact. If they will be preparing press materials for this manuscript, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact onepress@plos.org.

For any other questions or concerns, please email plosone@plos.org.

Thank you for submitting your work to PLOS ONE.

With kind regards,

PLOS ONE Editorial Office Staff

on behalf of

Professor Peter Lundberg

Academic Editor

PLOS ONE