INTRODUCTION
Our daily activities comprise of numerous actions involving tools. For these actions to be appropriate for object functions, quite refined neural processing must be invoked, along with retrieval of abstract knowledge on tools and their proper handling (Frey, Reference Frey2007; Vingerhoets, Reference Vingerhoets2014; see also Orban, Reference Orban2016; Reynaud, Lesourd, Navarro, & Osiurak, Reference Reynaud, Lesourd, Navarro and Osiurak2016). Indeed, the activation of relevant tool and action concepts, as well as all the necessary perceptual processing and visuomotor transformations preceding real activities with tools must be performed, or at least orchestrated, within a specialized network of brain regions, often referred to as the praxis representation network (PRN; see Frey, Reference Frey2008; Kroliczak & Frey, Reference Kroliczak and Frey2009). The temporo-parieto-frontal subdivisions of this network are commonly associated, respectively, with processing of object functionality, its incorporation into action plans, and conversion of the assembled information into deliberate motor acts (Andersen & Buneo, Reference Andersen and Buneo2002; Beurze, de Lange, Toni, & Medendorp, Reference Beurze, de Lange, Toni and Medendorp2007; Vingerhoets & Clauwaert, Reference Vingerhoets and Clauwaert2015). Yet, an often-neglected prerequisite of successful actions with tools is a proper grasp. The latter requires that the functional part of an object be distinguished from the remaining subdivisions of its structure, particularly its handle, whose spatial relation to the acting hand must be also appropriately computed. Put differently, even such simple tasks as grasping tools may already require quite sophisticated neural machinery (Przybylski & Kroliczak, Reference Przybylski and Kroliczak2017; see also McDowell, Holmes, Sunderland, & Schurmann, Reference McDowell, Holmes, Sunderland and Schurmann2018).
The majority of earlier studies concerning the role of networks underlying skilled actions with functional objects have focused almost exclusively on applying whole-brain, voxel-wise, that is, mass-univariate, analyses to investigating neural underpinnings of pantomimed and/or real actions with tools, with the grasp component typically missing (Hermsdorfer, Terlinden, Muhlau, Goldenberg, & Wohlschlager, Reference Hermsdorfer, Terlinden, Muhlau, Goldenberg and Wohlschlager2007; Johnson-Frey, Newman-Norlund, & Grafton, Reference Johnson-Frey, Newman-Norlund and Grafton2005; Valyear, Gallivan, McLean, & Culham, Reference Valyear, Gallivan, McLean and Culham2012; Vingerhoets, Vandekerckhove, Honore, Vandemaele, & Achten, Reference Vingerhoets, Vandekerckhove, Honore, Vandemaele and Achten2011; see also Brandi, Wohlschlager, Sorg, & Hermsdorfer, Reference Brandi, Wohlschlager, Sorg and Hermsdorfer2014; Kubiak & Kroliczak, Reference Kubiak and Kroliczak2016; Marangon, Kubiak, & Kroliczak, Reference Marangon, Kubiak and Kroliczak2016; Mizelle, Kelly, & Wheaton, Reference Mizelle, Kelly and Wheaton2013).
Notably, even in studies using univariate approaches with their subtraction of neural signals from control conditions, region of interest (ROI) analyses of percent signal change capitalize on neural activity averaged across a vector of voxels. Nevertheless, an increasing trend has been recently observed for analyzing neuroimaging data and interpreting their results in terms of patterns of activity (Fabbri, Stubbs, Cusack, & Culham, Reference Fabbri, Stubbs, Cusack and Culham2016; Gallivan, McLean, Valyear, Pettypiece, & Culham, Reference Gallivan, McLean, Valyear, Pettypiece and Culham2011; Gertz, Lingnau, & Fiehler, Reference Gertz, Lingnau and Fiehler2017; see also Shay, Chen, Garcea, & Mahon, Reference Shay, Chen, Garcea and Mahon2018). Such approaches provide additional explanatory power (Quadflieg et al., Reference Quadflieg, Etzel, Gazzola, Keysers, Schubert, Waiter and Macrae2011) and allow for characterizing neural representations as synchronized signal fluctuations (Haxby, Connolly, & Guntupalli, Reference Haxby, Connolly and Guntupalli2014; Norman, Polyn, Detre, & Haxby, Reference Norman, Polyn, Detre and Haxby2006). Namely, while in univariate analyses each voxel is treated independently, multivariate analyses typically reveal information encoded in a distributed manner across many voxels.
Indeed, this technique is known as multivoxel pattern analysis (MVPA) because signals from multiple spatial locations serve as inputs to (multivariate) models, whereas in mass-univariate tests information is analyzed for each voxel separately, and often combined into clusters of activity at the latest stages of analyses. Therefore, here we tested potential advantages of MVPA on data from a recent study conducted in our lab (Przybylski & Kroliczak, Reference Przybylski and Kroliczak2017), with the latter based mainly on the outcomes from the univariate, subtraction method.
Przybylski and Kroliczak (Reference Przybylski and Kroliczak2017) ran an fMRI study to investigate brain regions involved in processing visual and visuo-motor signals associated with preparation of functional grasps of tools. The stimuli were pictures of objects and the task was to plan and subsequently pantomime a proper grip. Planning functional grasps of tools was contrasted with planning grasps for control non-tool wooden objects (such as rods, sticks, etc.), with exemplars across both categories matched for their sizes and/or shape complexities.
The obtained results were such that regardless of the hand involved in planning functional grasps, the activity within the left-lateralized PRN complex was observed. This was rather surprising given that PRN is typically associated with the control of tool use gestures/actions, not the sole functional grip per se. Notably, such results are consistent with theoretical frameworks provided by ecological psychology (Gibson, Reference Gibson1977), wherein visuospatial processing of an object, affordances such as grasp, and action execution are largely inseparable (cf. Osiurak, Rossetti, & Badets, Reference Osiurak, Rossetti and Badets2017).
Using Przybylski and Kroliczak’s (Reference Przybylski and Kroliczak2017) data and MVPA, here we tested where and how neural response patterns encode grasp planning information, and whether multivariate analyses corroborate outcomes from the univariate subtraction technique previously used. After all, it is often suggested that MVPA provides more “fine-grained” outcomes (Kriegeskorte, Goebel, & Bandettini, Reference Kriegeskorte, Goebel and Bandettini2006). From a broader perspective, especially in the light of the forthcoming brain-machine interface possibilities, it is essential to study preparatory neural mechanisms preceding common intentional actions. In this fast-developing field of research and its applications, an advantage of decoding-based approaches over measuring differences between means seems indisputable (Andersen, Musallam, & Pesaran, Reference Andersen, Musallam and Pesaran2004; Haynes & Rees, Reference Haynes and Rees2006).
METHODS
Participants
Data from 20 healthy participants identical to those from a study by Przybylski and Kroliczak (Reference Przybylski and Kroliczak2017) were used in this report: they were all native Polish-speaking students (10 females; age 19 to 24 years; mean=22.7; SD=1.6), right-handed (Edinburgh Handedness Inventory index: M=92.9; SD=13.9; Oldfield, Reference Oldfield1971), with normal or corrected to normal visual acuity. Written informed consent was obtained from each individual. After the study participants were compensated financially for their time and effort, and were debriefed. All protocols and procedures, which conformed to the principles of the 2013 WMA Helsinki Declaration, were approved by the Bio-Ethics Committee at Poznan University of Medical Sciences.
Stimuli
Photographs of 12 tools were used as stimuli. Each object was shown in the foreshortened perspective, that is, as if the viewer was standing in the front of the table on which an object was laid down. The experimental stimuli were chosen from kitchen, workshop and garden tools (see examples in Figure 1A, left panel). Control objects were pictures of 12 wooden pieces, such as parts of the branches or man-made wooden sticks (see Figure 1A, right panel). Although all the stimuli were well-known common objects, before the study participants were familiarized with them and with all the tasks to be performed during experiment proper. Graspable parts of tools and control stimuli were matched for size and/or complexity and were not expected to contribute to the observed variance as a confounding factor. Each object was presented in one of three different views (orientations), which within each category of objects assured no differences in task difficulty for the grasping hand (left or right). Because in their original report Przybylski and Kroliczak (Reference Przybylski and Kroliczak2017) demonstrated that the greater engagement of the parieto-frontal networks for tools was observed regardless of object orientation, the orientation factor was omitted in the current analyses.
Examples of stimuli and an experimental paradigm (adapted with permission of Cambridge University Press from Przybylski & Kroliczak, Reference Przybylski and Kroliczak2017). (A) The two classes of stimuli and their orientations. The functional items were simple tools (left panel) and control stimuli were wooden pieces and branches (right panel). Each object was presented at three different orientations: 0°, 125°, and 225°. (B) A schematic of an event-related paradigm used in this study. The grasp-planning phase started with stimulus onset and lasted for a variable interval of 3, 4, or 5s. Subsequently, participants pantomimed the preplanned grasp for the duration of the “Grasp” cue (1.5s). A trial was pseudo-randomly followed by either a variable inter-trial interval (ITI of 2.5, 3.5, or 4.5s) or a period of 10-s rest.
Procedure
On two separate (typically consecutive) days, volunteers participated in two scanning sessions wherein they performed the task with their right (dominant) or left (non-dominant) hands, with the order of the tested hand counterbalanced. In an event-related paradigm, a varying ISI (0.0, 0.25, 0.5, or 0.75 s) was introduced at the beginning of each trial to jitter the stimulus onset with respect to the acquisition of a functional volume, therefore, artificially improving temporal resolution of the fMRI scans (Miezin, Maccotta, Ollinger, Petersen, & Buckner, Reference Miezin, Maccotta, Ollinger, Petersen and Buckner2000). Subsequently, the stimulus picture was displayed on the screen for 1.5 s, which coincided with the beginning of the grasp-planning phase, lasting for additional 1.5, 2.5, or 3.5 s following stimulus offset (3, 4, or 5 s in total).
Then, the “Grasp” cue (green circle) was shown for 1.5 s and participants’ task was to pantomime the preplanned grasp, either functional or control, for tools or non-tools, respectively. A given trial concluded with an inter-trial interval of 2.5, 3.5, or 4.5 s, or a 10-s rest period. The protocol schematic is illustrated in Figure 1B. Grasp-planning activity in both conditions was modeled as the 3-s period beginning with the onset of the stimulus picture (displayed for 1.5 s) and lasting through the end of the shortest (1.5-s) delay interval (an approach inspired by Kroliczak & Frey, Reference Kroliczak and Frey2009).
Data Acquisition
Siemens 3 Tesla MAGNETOM Trio MRI scanner in the Laboratory of Brain Imaging (http://lobi.nencki.gov.pl) in Warsaw, Poland, was used to perform imaging with a 32-channel head coil. For better slice prescriptions, before functional runs, Auto Align Scout and True FISP sequences were applied. T2 *-weighted gradient echo sequence was used to acquire blood-oxygen-level-dependent (BOLD) echo-planar images: time repetition (TR)=2000 ms; time to echo (TE)=30 ms; flip angle (FA)=90°; 64 × 64 matrix; field of view (FOV)=200 mm; 34 contiguous axial slices; 3.1-mm isotropic voxels (each run consisted of 145 such volumes).
We also acquired T1-weighted structural images using magnetization prepared rapid gradient echo (MP-RAGE) pulse sequence: TR=2540 ms; TE=3.32 ms; inversion time (TI)=1200 ms; FA=7°; 256 × 256 voxel-matrix; FOV=256 mm; 176 contiguous axial slices; 1.0-mm isotropic voxels. Additionally, to improve registration between native space (functional echo planar images) and anatomical T1-weighted scans, T2-weighted structural images were collected: TR=3200 ms; TE=402 ms; FA=120°; 512 × 512 voxel-matrix; FOV=256 mm; 176 contiguous axial slices; 0.5 × 0.5 × 1.0 mm non-isotropic voxels. We converted raw data to NIfTI-1 format using MRI-Convert software (http://lcni.uoregon.edu/downloads/mriconvert).
Preprocessing
Functional data were preprocessed using FSL (FMRIB’s Software Library) (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki), with the following procedures: brain extraction (BET), motion correction (MCFLIRT), and high-pass filtering (σ=50.0 s). No spatial smoothing was used. Because the main task always consisted of five functional runs per hand, and the analysis was performed separately for each hand, as well as compared across subjects and/or hands, all runs were co-registered (using FLIRT) to common space, i.e., the middle volume of the first functional run. The very same reference (i.e., middle) volume was also spatially normalized (using matrices from the registration of the first run) to the MNI 152 2-mm standard space trough initial, high-resolution T2-weighted and standard T1-weighted structural images. All the operations were performed with either FSL FEAT (FMRI Expert Analysis Tool) v5.0.9 (Jenkinson, Beckmann, Behrens, Woolrich, & Smith, Reference Jenkinson, Beckmann, Behrens, Woolrich and Smith2012) or its subcomponents and commands. The same anatomical data and processing procedures as in Przybylski and Kroliczak (Reference Przybylski and Kroliczak2017) were used for spatial registration of the structural (T2- and T1-weighted) images.
MVPA
Because a fast event-related design was used in this experiment, parameter estimates (PEs) for each trial had to be calculated first. To this end, a general linear model (GLM) was created separately for each trial using the least squares-separate approach (LS-S, for details see: Mumford, Turner, Ashby, & Poldrack, Reference Mumford, Turner, Ashby and Poldrack2012). Each GLM was created for every trial within each run, every time with two explanatory variables: the first for the trial of interest, and the second for all the remaining other trials within a given run. Specifically, as there were 12 trials per condition during one functional run (i.e., 24 trials in total per run) then 24 GLMs were created to obtain one PE for each trial (24 × 5 runs, which gives 120 matrices of beta values altogether). Motion parameters, cerebrospinal fluid, and white-matter time series were used as additional regressors. PEs were not normalized before the multivariate analysis. (We would have considered normalizing PEs if the outcomes just replicated the univariate results; cf. Misaki, Kim, Bandettini, & Kriegeskorte, Reference Misaki, Kim, Bandettini and Kriegeskorte2010). The two classified conditions were: (1) planning functional grasps of tools; and (2) planning control grasps of non-tool objects. In other words, planning of tool-oriented grasps was always classified against planning grasps of non-tools.
Subsequently, the n-fold cross-validation method was used, where n was the number of runs (an approach often called leave-one-run-out method). We evaluated classifier performance using the accuracy metric. Because the number of trials per condition was equal, there was no risk of the accuracy paradox. As there were 12 trials per condition for one run, and five runs in total, each fold of the validation consisted of 96 PEs (12 × 2 × 4; trials × conditions × runs) for training the classifier and 24 (12 × 2 × 1; trials × conditions × run) PEs to test the accuracy of the classification. This operation was performed five times, and the classification accuracies were averaged (with arithmetic mean). This way the single accuracy score was obtained per each hand for each participant in a range of 0.0 to 1.0 (i.e., 0% to 100%). MVPA classification was performed with the support vector machine (SVM)-based classifier (linear kernel, C parameter fixed at 1.0) implemented in Python’s package scikit-learn (Abraham et al., Reference Abraham, Pedregosa, Eickenberg, Gervais, Mueller, Kossaifi and Varoquaux2014; http://scikit-learn.org/stable), via the nilearn module (http://nilearn.github.io).
ROI Selection
Based on previous research, 12 left-hemispheric ROIs were selected: (1) human intraparietal area 1 (hIP1); (2) dorsal intraparietal sulcus, anterior division (DIPSA); (3) dorsal intraparietal sulcus, medial division (DIPSM); (4, 5, and 6) supramarginal gyrus (its cytoarchitectonic subdivisions: PF, PFm, and PFt); (7) putative human homologue of the anterior intraparietal area (phAIP/aIPS); (8) superior parieto-occipital cortex (SPOC); (9) ventral premotor cortex (PMv); (10) dorsal premotor cortex (PMd); (11) caudal middle temporal gyrus (cMTG), and (12) rostral middle frontal gyrus (rMFG). Additionally, one control (13th) region was included in the analysis, namely, the right lateral ventricle (rLV). For more detailed information on the sources of all ROIs, see Table 1. The borders of the ROIs mapped to the inflated and flat brain surfaces are shown in Figure 2.
ROIs shown on: (A) the flattened brain surface, (B) the partly inflated lateral, and (C) dorsal surface of the left hemisphere. hIP1, human intraparietal area 1; DIPSA, dorsal intraparietal sulcus, anterior division; DIPSM, dorsal intraparietal sulcus, medial division; PF, PFm, PFt, supramarginal gyrus subdivisions; phAIP, putative human (homologue of the) anterior intraparietal area; SPOC, superior parieto-occipital cortex; PMv, ventral premotor cortex; PMd, dorsal premotor cortex; cMTG, caudal middle temporal gyrus; rMFG, rostral middle frontal gyrus; rLV, right lateral ventricle. The sizes of the ROIs in individual subjects’ native spaces varied depending on structural differences.
Regions of interest used in the analysis
Note. The number of voxels was calculated for ROIs in MNI 152 brain template. The sizes of the ROIs in individual subjects’ native spaces varied depending on structural differences. See Figure 2 for the visualization on the brain surfaces.
hIP1=human intraparietal area 1; DIPSA=dorsal intraparietal sulcus, anterior division; DIPSM=dorsal intraparietal sulcus, medial division; PF, PFm, PFt=supramarginal gyrus (SMG) subdivisions; phAIP=putative human homologue of the anterior intraparietal; SPOC=superior parieto-occipital cortex; PMv=ventral premotor cortex; PMd=dorsal premotor cortex; cMTG=caudal middle temporal gyrus; rMFG=rostral middle frontal gyrus; rLV=right lateral ventricle.
Before MVPA, ROIs were transformed from standard space (MNI152 2mm) to individual subjects’ native spaces using FLIRT (with transformation matrices obtained during a preprocessing step, from the middle volume of the first functional run, through T2- and T1-weighted images, see the Preprocessing section). The values from voxels (beta weights) within the range of respective masks were used for training and testing the classifier.
ROI t Tests
For the following t test statistical procedures classification scores were averaged (with arithmetic mean) across validation folds; hence there were 20 (number of participants) values per ROI. To assess statistical significance of the decoding within the functional ROIs, a one-sample t test was run across participants (separately for each hand) with respect to 50% chance level. This was followed by Bonferroni correction for multiple comparisons (25 tests in total, i.e., 2 tests per hand for 12 functional ROIs and 1 for the control site).
Furthermore, to provide more exact information about the actual contribution of the particular region, the performance of each hand within each ROI was contrasted with the corresponding results for our control ROI (rLV), whose classification accuracy was averaged across hands before the t tests. The comparison was achieved by running 24 Bonferroni-corrected paired t tests (12 for the dominant hand and 12 more for the left hand). Finally, we opted for determining which of the ROIs differed in terms of classification accuracies depending on whether the action was planned for the dominant or non-dominant hand. Since the accuracy scores within binary independent variable were tested, p-values were not adjusted for multiple comparisons in this case.
Before t tests, we conducted two-sigma outlier detection and normality assessment with Shapiro-Wilk test. All statistical procedures within this section were carried out with Python’s SciPy module, v. 0.19.0 (Oliphant, Reference Oliphant2007), with corrections for multiple tests performed with StatsModels, v. 0.6.1 (Seabold & Perktold, Reference Seabold and Perktold2010).
Searchlight Analysis
In addition to ROI analyses, a whole-brain searchlight procedure was also applied to reveal further nodes of the network (brain areas specific to the planning of functional grasps of tools) and/or to validate the results obtained with the ROIs from earlier analyses. A radius of 5.6 mm around the voxel of interest was the spherical region for each searchlight classification (with SVM classifier and C parameter at 1.0). Cross-validation was the same as for ROI decoding (n-fold cross validation). The resulting five scores (5 × k matrix) per participant and per hand were averaged (arithmetic mean) to create a single vector of variables, of the length k (where k was the number of voxels in particular subject’s native space). The resulting maps (one for each of the 20 participants) were then spatially transformed with registration matrices from the first functional run to the MNI152 2-mm standard atlas.
A second-level searchlight analysis was performed by: (1) combining the first-level brain maps (individual participant’s results) with one sample t test (null hypothesis being that the classification is at 50% chance level), for the right and left hand separately; (2) a direct comparison between the right and left hand, with a paired t test. The outcomes of both contrasts were subsequently corrected for multiple comparisons with the Bonferroni method.
RESULTS
ROI t Test Results
Significant classifications against chance level for each ROI and for each hand separately are depicted in Figure 3 as red asterisks. Except for rLV, the results from all ROIs were significantly different from chance (α=0.05). Notably, significance levels achieved for the studied regions ranged from 1*10−3 to 5*10−10. The lowest p-values were obtained within PMd, DIPSM, and phAIP. The least probable effects were observed in the PMv, PFm, cMTG, rMFG, and hIP1. Using nonparametric permutation tests (Etzel, Reference Etzel2017) rather than t tests yielded qualitatively similar results.
The results of the ROI analyses. Green bars (left ones) represent mean accuracy scores for the left-hand, and red ones for the right hand (averaged, with arithmetic mean, across folds and subjects). Red asterisks (upper set) indicate statistically significant differences with respect to chance level (at 50%) as measured with two-tailed one-sample t tests. The significance against the control region (rLV, the last bar on the right with thick blue borders) is denoted by blue asterisks (lower set), the results of pairwise t test comparisons. Error bars are standard errors of measurement. Thick, solid, black horizontal line represents the chance level. Abbreviations as in Figure 2.
When praxis-related ROIs were compared to the reference area, all but one (rMFG) showed statistically significant differences at least at p<.05 (regardless of the hand). These results are marked with blue asterisks on the graph in Figure 3. All t test comparisons within ROIs (between hands) turned out to be statistically non-significant.
Searchlight Analysis Results
This analysis revealed that the information required to decode planning functional grasps of tools is processed mainly in the left cerebral hemisphere (regardless of the hand). Yet, when participants were asked to plan grasps with their dominant hand, the pattern of statistically significant voxels included: (1) PMd, extending to the superior frontal gyrus (SFG); (2) SPOC vicinity, extending to the dorsal IPS; (3) small clusters within SMG (PFt, PF); as well as additional locations in (4) area V3, LO1 and the ventral lateral occipital complex (vLOC), and (5) the supplementary motor area (SMA). These effects are visualized in Figure 4A. Right-hemispheric voxels were located mainly in: (1) vLOC; (2) antero-dorsal precuneus; (3) dorsal IPS (including DIPSA/DIPSM); (4) PF; and (5) three small clusters in superior frontal cortices (Figure 4B).
The results of searchlight analysis for functional grasp planning with the right hand. Specifically, this figure depicts statistically significant decoding accuracies of planning functional grasps of tools versus non-tools. The unit with the red-yellow bar is a Z-score (standardized t-score, from one-sample t test across subjects against 50% prior chance decoding accuracy, corrected for multiple comparisons with Bonferroni method at p<.05) and thresholded at 3.1 value. Names of the ROI borders mapped to the left hemisphere surface can be found in Figure 2. Right-hemisphere ROIs are shown here for convenience. These are as follows: SPOC (blue; left upmost), DIPSM, DIPSA, phAIP (white, top to bottom; between SPOC and PF complex); PFm, PF and PFt (green, left to right; bottom cluster of borders).
For functional grasp planning with the left-hand, the left-hemisphere statistically significant patterns were found in: (1) PMd; (2) phAIP/aIPS; (3) DIPSA, DIPSM; (4) small clusters in PFt and PFm; and (5) vLOC. For activity patterns see Figure 5A. In the right hemisphere, the most important clusters were located in: (1) DIPSA and DIPSM; (2) PMd; (3) PFt; and (4) vLOC (see Figure 5B). A direct comparison between maps for the two hands (treated as dependent groups) with paired t test revealed no statistically significant results.
The results of searchlight analysis for functional grasp planning with the left hand. As before (Figure 4), this figure depicts statistically significant decoding accuracies of planning functional grasps of tools versus non-tools. The units are Z-scores from one-sample t test against 50% chance level, corrected for multiple comparisons with Bonferroni method at p<.05 and thresholded at 3.1 value. For ROI names see Figure 2 (left hemisphere) and Figure 4 (right hemisphere).
Figure 2, Figure 4, as well as Figure 5 were generated using the Caret software (https://www.nitrc.org/projects/caret/; Van Essen et al., Reference Van Essen, Drury, Dickson, Harwell, Hanlon and Anderson2001). To enable comparisons of our results with future studies that will use the Connectome Workbench (a potential standard) in data visualization (https://www.humanconnectome.org/software/connectome-workbench; Marcus et al., Reference Marcus, Harwell, Olsen, Hodge, Glasser, Prior and Van Essen2011), in Figure 6, we also mapped our searchlight patterns onto its inflated (midthickness) and flat surfaces, emphasizing the associated areas (of the 180 identified and delineated parcellations; Glasser et al., Reference Glasser, Coalson, Robinson, Hacker, Harwell, Yacoub and Van Essen2016). Only the surfaces of the left hemisphere, typically associated with praxis skills, were used to map these results.
MVPA results displayed on flat and inflated (midthickness) surfaces from the Connectome Workbench (v. 1.2.3.) visualization software (Glasser et al., Reference Glasser, Coalson, Robinson, Hacker, Harwell, Yacoub and Van Essen2016) of the Human Connectome Project (HCP). The borders of 180 areas from the HCP’s parcellation are also shown, and the critical ones provided with their labels (which are described in details in Supplemental Materials for Glasser et al., Reference Glasser, Coalson, Robinson, Hacker, Harwell, Yacoub and Van Essen2016).
DISCUSSION
The outcomes obtained with MVPA are in substantial agreement with earlier studies demonstrating that the control of praxic skills is primarily left lateralized and involves the occipito-temporal, as well as parieto-frontal subdivisions of PRN (Frey, Reference Frey2008; see also Kristensen, Garcea, Mahon, & Almeida, Reference Kristensen, Garcea, Mahon and Almeida2016; Przybylski & Kroliczak, Reference Przybylski and Kroliczak2017; Vingerhoets & Clauwaert, Reference Vingerhoets and Clauwaert2015). This network of areas, with SMG as its central node, is typically associated with representing manipulation knowledge (Buxbaum, Reference Buxbaum2001; Buxbaum & Kalenine, Reference Buxbaum and Kalenine2010; cf. Goldenberg, Reference Goldenberg2017; Lesourd, Osiurak, Navarro, & Reynaud, Reference Lesourd, Osiurak, Navarro and Reynaud2017) or mechanical knowledge (Goldenberg & Hagmann, Reference Goldenberg and Hagmann1998; Osiurak et al., Reference Osiurak, Jarry, Allain, Aubin, Etcharry-Bouyx, Richard and Le Gall2009; Osiurak, Jarry, Lesourd, Baumard, & Le Gall, Reference Osiurak, Jarry, Lesourd, Baumard and Le Gall2013; see also Lesourd, Budriesi, Osiurak, Nichelli, & Bartolo, Reference Lesourd, Budriesi, Osiurak, Nichelli and Bartolo2017) linked to object functionality, and its integration into action plans for programming of deliberate motor responses (Bernier, Cieslak, & Grafton, Reference Bernier, Cieslak and Grafton2012; Fabbri et al., Reference Fabbri, Stubbs, Cusack and Culham2016; Gallivan et al., Reference Gallivan, McLean, Valyear, Pettypiece and Culham2011).
The observed left-brain bias, revealed regardless of the used hand, indicates that PRN contains hand-independent representations of praxis skills (Kroliczak, Piper, & Frey, Reference Kroliczak, Piper and Frey2016). Nevertheless, MVPA exposed greater contributions from aIPS, an area linked to the control of grasping movements (e.g., Cavina-Pratesi, Goodale, & Culham, Reference Cavina-Pratesi, Goodale and Culham2007; Kroliczak, McAdam, Quinlan, & Culham, Reference Kroliczak, McAdam, Quinlan and Culham2008), and vLOC, associated with encoding visual characteristics of tools, their concepts, and the retrieval of their function-relevant features (e.g., Bracci, Cavina-Pratesi, Ietswaart, Caramazza, & Peelen, Reference Bracci, Cavina-Pratesi, Ietswaart, Caramazza and Peelen2012; Garcea & Mahon, Reference Garcea and Mahon2014; Nastase et al., Reference Nastase, Connolly, Oosterhof, Halchenko, Guntupalli, Visconti di Oleggio Castello and Haxby2017; Oosterhof, Wiggett, Diedrichsen, Tipper, & Downing, Reference Oosterhof, Wiggett, Diedrichsen, Tipper and Downing2010; see also Peelen et al., Reference Peelen, Bracci, Lu, He, Caramazza and Bi2013; Wurm & Lingnau, Reference Wurm and Lingnau2015).
Core Areas for Planning Functional Grasps
Praxis-related literature emphasizes the importance of left aSMG as the core region responsible for the integration of signals from posterior SMG, IPS, as well as from the tempo-occipital cortices (e.g., Orban & Caruana, Reference Orban and Caruana2014; Peeters, Rizzolatti, & Orban, Reference Peeters, Rizzolatti and Orban2013), thus playing a critical role in motor cognition. However, MVPA results obtained here for functional grasp planning highlight the role of aIPS and vLOC, the two areas that likely provide prerequisite inputs to SMG for performing actions with tools. Indeed, these outcomes are consistent with a notion that manipulation knowledge is supported by perceptual representations in the ventral stream (e.g., Almeida, Fintzi, & Mahon, Reference Almeida, Fintzi and Mahon2013; Mahon, Kumar, & Almeida, Reference Mahon, Kumar and Almeida2013; see also Gallivan, Cant, Goodale, & Flanagan, Reference Gallivan, Cant, Goodale and Flanagan2014), and their projections to SMG interact with manipulation-based action representations stored or elaborated there (Buxbaum, Reference Buxbaum2017; cf. Osiurak & Badets, Reference Osiurak and Badets2016; Reynaud et al., Reference Reynaud, Lesourd, Navarro and Osiurak2016). In this framework, vLOC would also contribute to object parsing, including identification of functional and graspable subdivisions, thus allowing aIPS to concentrate almost exclusively on processing object handle(s)/graspable part(s).
It is of note, however, that both our earlier study (Przybylski & Kroliczak, Reference Przybylski and Kroliczak2017) and the current outcomes indicate that the traditionally defined cMTG (e.g., Kroliczak & Frey, Reference Kroliczak and Frey2009; see also Garcea, Kristensen, Almeida, & Mahon, Reference Garcea, Kristensen, Almeida and Mahon2016; Vingerhoets, Nys, Honore, Vandekerckhove, & Vandemaele, Reference Vingerhoets, Nys, Honore, Vandekerckhove and Vandemaele2013) was barely involved in the planning of functional grasps. Indeed, regardless of the hand, the multivoxel patterns of brain activity were observed only in the nearby occipital cortex. Yet, overlapping clusters in very similar locations have been shown to respond both to tools and hands (Bracci, Ritchie, & de Beeck, Reference Bracci, Ritchie and de Beeck2017; Striem-Amit, Vannuscorps, & Caramazza, Reference Striem-Amit, Vannuscorps and Caramazza2017) and were also reported engaged in some aspects of matching of hand posture to object orientation (Vingerhoets et al., Reference Vingerhoets, Nys, Honore, Vandekerckhove and Vandemaele2013), a kind of processing that could be considered a prerequisite of further interactions with tools. All things considered, planning functional grasps of tools requires the reasoning about object properties associated with activity within vLOC and SMG subdivisions putatively interacting with it, as well as integrating signals from aIPS (Chao & Martin, Reference Chao and Martin1999; Grill-Spector & Weiner, Reference Grill-Spector and Weiner2014; Orban & Caruana, Reference Orban and Caruana2014).
Although MVPA reveals some local differences in activity patterns observed in areas PFt, PF, and PFm between hands (see Figure 7A–D), a direct between-hands contrast convincingly demonstrates that the neural underpinnings of higher-order actions should be virtually identical. The results of ROI analyses (e.g., PFt, PF, and PFm ROIs in Figure 3) are consistent with this notion. Nevertheless, in individuals with somewhat atypically represented praxic functions this relationship may change and the extent of neural activity accompanying the use of the better-trained or the dominant hand may show closer affinity to the degree and direction of handedness (e.g., Dassonville, Zhu, Uurbil, Kim, & Ashe, Reference Dassonville, Zhu, Uurbil, Kim and Ashe1997; cf. Drager et al., Reference Drager, Jansen, Bruchmann, Forster, Pleger, Zwitserlood and Knecht2004; Biduła, Przybylski, Pawlak, & Kroliczak, Reference Biduła, Przybylski, Pawlak and Kroliczak2017; see also Kroliczak, Buchwald, Potok, & Przybylski, Reference Kroliczak, Buchwald, Potok and Przybylski2018).
Subdivisions of left inferior parietal lobule and their involvement in planning functional grasps as revealed by traditional (contrast) analyses and MVPA. The outcomes are shown on flat brain surfaces using Caret software. The outcomes for the right hand: (A) activity patterns adopted from a study by Przybylski and Kroliczak (Reference Przybylski and Kroliczak2017) and (B) the multivoxel patterns of brain activity from the searchlight analysis (MVPA). The outcomes for the left hand: (C) activity patterns from a study by Przybylski and Kroliczak (Reference Przybylski and Kroliczak2017) and (D) the multivoxel patterns of brain activity from the current analysis (MVPA).
In sum, the current outcomes obtained with MVPA are consistent with the notion that transformations of object-related information into goal-directed actions are computed across a much broader range of brain areas or regions (Gallivan & Culham, Reference Gallivan and Culham2015), even if actions are limited to planning functional grasps. Nevertheless, it should be mentioned that the task used here, that is, participants planned and executed pantomimed functional grasps, does not necessarily invoke typical visuomotor (i.e., dorsal stream; Goodale & Milner, Reference Goodale and Milner1992) mechanisms for reach-to-grasp actions directed at objects under guidance of visual feedback in real time (Kroliczak, Cavina-Pratesi, Goodman, & Culham, Reference Kroliczak, Cavina-Pratesi, Goodman and Culham2007; Gallivan, McLean, Valyear, & Culham, Reference Gallivan, McLean, Valyear and Culham2013; Freud et al., Reference Freud, Macdonald, Chen, Quinlan, Goodale and Culham2018).
Instead, a greater contribution from the ventral stream would be expected (Goodale, Jakobson, & Keillor, Reference Goodale, Jakobson and Keillor1994; Rossit et al., Reference Rossit, Malhotra, Muir, Reeves, Duncan and Harvey2011), particularly because pantomimed grasps resembled manual estimations (Kroliczak, Westwood, & Goodale, Reference Kroliczak, Westwood and Goodale2006), although taking into account the requisite hand rotations. Yet, the stimuli were functional objects, and adaptive behavior in their presence requires contributions of the dorsal and ventral streams (Goodale, Kroliczak, & Westwood, Reference Goodale, Kroliczak and Westwood2005), with numerous functional subdivisions of the former. The remaining clusters will be discussed next.
Remaining regions
MVPA also revealed the involvement of several further regions in transforming visual information into plans for functional grasps of tools. Although the greater engagement of these auxiliary areas in the more demanding actions incorporating tools is not surprising, it is unlikely that their contribution to task performance is tool specific.
Neural activity within IPS (namely phAIP, DIPSA, and/or DIPSM) is consistent with an idea that IPS does not code the detailed movement kinematics but rather the to-be-obtained action goals (Hamilton & Grafton, Reference Hamilton and Grafton2008). Yet another reason for its involvement is the fact that a tool’s handle affords “a grasp” more noticeably than a corresponding part in the control (non-tool) object (Vingerhoets, Reference Vingerhoets2014).
Another vicinity revealed with MVPA was PMd, which is often associated with neural underpinnings of the preparation and control of visuomotor processes involving task-response associations and action sequencing (Cross, Hamilton, Cohen, & Grafton, Reference Cross, Hamilton, Cohen and Grafton2017). PMd also belongs to the dorso-dorsal stream, composed of the V6/V6A–dorso-medial intraparietal–PMd circuit (Rizzolatti & Matelli, Reference Rizzolatti and Matelli2003). A contribution from all these nodes was revealed by our analysis. Although it has been speculated that this circuit is involved in on-line control of actions (Binkofski & Buxbaum, Reference Binkofski and Buxbaum2013; see also Milner & Goodale, Reference Milner and Goodale2008), the on-line component does not extend here to the actually performed manual tasks but a mere preparation for them. This outcome is consistent with the postulated contribution of the dorso-dorsal stream in processing variable affordances, that is temporary and/or transient object characteristics such as changes in their orientation, size or shape, and the required adjustments of grip/arm rotation for the task at hand (Kroliczak et al., Reference Kroliczak, McAdam, Quinlan and Culham2008; Sakreida et al., Reference Sakreida, Effnert, Thill, Menz, Jirak, Eickhoff and Binkofski2016).
Notably, while the PMd contribution was substantial, the SPOC vicinity was underrepresented here. Yet, the stimuli were not within reach (they were pictures on the screen) and, again, the mechanisms involved in pantomimed grasp planning (whether functional or not) can be different from preparation for real reaching and grasping (Gallivan, Cavina-Pratesi, & Culham, Reference Gallivan, Cavina-Pratesi and Culham2009). Furthermore, these objects were not presented in lower visual fields, which are preferentially processed for hand actions by SPOC (Rossit, McAdam, McLean, Goodale, & Culham, Reference Rossit, McAdam, McLean, Goodale and Culham2013). Nevertheless, our results are quite consistent with recent outcomes revealing grater connectivity between cMTG/LOC and aIPS (Chen, Snow, Culham, & Goodale, Reference Chen, Snow, Culham and Goodale2018), with some contribution from the superior parietal lobule (SPL) in a contrast of tools versus non-tools (whether elongated or stubby). As in their study, the SPL activity we observed here, regardless of the hand used, extended toward SPOC.
SMA, typically associated with the control of visually guided movements (Picard & Strick, Reference Picard and Strick2003) also contributed to planning functional grasps of tools. The patterns of neural activity elicited by such skilled motor acts (including finger sequencing) are often decoded most accurately within this area (Wiestler & Diedrichsen, Reference Wiestler and Diedrichsen2013). Although grasp performance was pantomimed (with no further tool-use actions), the task nevertheless involved internally generated movement plans which also require SMA (Elsinger, Harrington, & Rao, Reference Elsinger, Harrington and Rao2006), whose inputs are critical in the process of predictive motor planning (cf. Makoshi, Kroliczak, & van Donkelaar, Reference Makoshi, Kroliczak and van Donkelaar2011).
Conclusions
The engagement of the temporo/occipital-parieto-frontal cortices in processing functional grasps of tools revealed by MVPA is largely consistent with the outcomes from traditional subtraction analyses. Nevertheless, the multivoxel patterns of brain activity, which are obtained with different theoretical and algorithmic framework from these that underlie subtraction contrasts, put greater emphasis on grasp coding in aIPS, the occipito-temporal encoding of visual/structural and functional features of tools, as well as on sensorimotor, dorso-dorsal preparatory processes, rather than on the integrative role of aSMG.
The ventral occipital regions may provide auxiliary inputs for grip formation and the retrieval of critical, tool-related knowledge on their invariant features, that is, stable affordances, that underlie reasoning about tool properties based on manipulation and mechanical knowledge most likely stored in left SMG. All in all, MVPA is a reliable technique that may greatly enrich our understanding of information processing preceding any functional interactions with tools.
ACKNOWLEDGMENTS
The authors declare no conflict of interest. This work was supported by National Science Centre (Narodowe Centrum Nauki, NCN) grant Maestro 2011/02/A/HS6/00174 to Gregory Kroliczak. During all stages of this work M.B., Ł.P., and G.K. were also supported by the Maestro grant. The equipment used for data acquisition and analyses was funded by the Ministry of Science and Higher Education (Ministerstwo Nauki i Szkolnictwa Wyższego, MNiSW) grant 6168/IA/128/2012 to Gregory Króliczak. The CePT research infrastructure was purchased with funds from the European Regional Development Fund as part of the Innovative Economy Operational Programme, 2007–2013. Author contributions: This project was conceptualized by G.K., Ł.P., and M.B. Data were collected by Ł.P. and G.K., analyzed by M.B., and interpreted by M.B, G.K., and Ł.P. The manuscript was written by M.B. and G.K. The authors thank Guy Orban and David Van Essen for introducing us to new methods of parcellation and visualization of neuroimaging data.
