The neuropsychological syndrome “simultanagnosia” is characterized by the inability to integrate local elements into a global entity. This deficit in Gestalt perception is mainly apparent for novel global structures administered in clinical tests or unfamiliar visual scenes. Recognition of familiar complex objects or well-known visual scenes is often unaffected. Recent neuroimaging studies and reports from simultanagnosia patients suggest a crucial involvement of temporoparietal brain areas in processing of hierarchically organized visual material. In this study, we investigated the specific role of the TPJ in Gestalt perception. On the basis of perceptual characteristics known from simultanagnosia, we hypothesized that TPJ is dominantly involved in processing of novel object arrangements. To answer this question, we performed a learning study with hierarchical stimuli and tested behavioral and neuronal characteristics of Gestalt perception pre- and posttraining. The study included 16 psychophysical training sessions and two neuroimaging sessions. Participants improved their behavioral performance for trained global stimuli and showed limited transfer to untrained global material. We found significant training dependent neuronal signal modulations in anterior right hemispheric TPJ regions. These activation changes were specific to trained global stimuli, whereas no systematic neuronal response changes were observed for recognition of untrained global stimuli, local elements and regular objects that served as control stimuli. In line with perceptual characteristics in simultanagnosia, the results argue for an involvement of TPJ in processing of novel global structures. We discuss the signal modulations in the context of a more efficient or different neuronal strategy to process familiar global stimuli.
A basic feature of visual perception and spatial orienting is the grouping of single elements into a superior global entity or so-called “Gestalt” (Koffka, 1935; Wertheimer, 1923). The relevance of such visual top–down organization is emphasized through a neuropsychological deficit termed “simultanagnosia” (Wolpert, 1924; Bálint, 1909), that is, the inability to specifically recognize a global stimulus arrangement. Patients suffering from this impairment are able to perceive single objects but cannot recognize meaningful configurations of several elements or objects. Evidence from lesion patterns in neurological patients with simultanagnosia as well as functional neuroimaging studies in healthy participants suggests a crucial role of bilateral TPJ areas in Gestalt perception (Rennig, Bilalić, Huberle, Karnath, & Himmelbach, 2013; Huberle & Karnath, 2012; Dalrymple, Bischof, Cameron, Barton, & Kingstone, 2010; Huberle, Driver, & Karnath, 2010; Himmelbach, Erb, Klockgether, Moskau, & Karnath, 2009; Clavagnier, Fruhmann Berger, Klockgether, Moskau, & Karnath, 2006; Huberle & Karnath, 2006; Weissman & Woldorff, 2005; Yamaguchi, Yamagata, & Kobayashi, 2000; Friedman-Hill, Robertson, & Treisman, 1995). These findings about neuronal correlates of visual Gestalt perception require an extension of our established knowledge about neuronal processing of visual information along the ventral visual stream (Grill-Spector et al., 1999; Goodale & Milner, 1992). Various studies have suggested a hierarchical axis of object processing along the ventral visual stream from local features encoded in early visual to global representations emerging in higher object-sensitive areas (Lerner, Hendler, Ben-Bashat, Harel, & Malach, 2001; Fink et al., 1996, 1997; Malach et al., 1995). From this perspective, it is striking that the hierarchical processing of complex arrangements and objects—the Gestalt of a rich visual scene—is not just another function of inferior temporal cortex but obviously driven by a distinct, much more dorsally located region, namely area TPJ.
Although we know about its general involvement, the specific contribution of (bilateral) TPJ areas to global Gestalt perception is fairly unknown. A possible function could be the assembling of novel stimulus configurations. Although familiar objects and object arrangements are processed without significant TPJ contribution, TPJ regions are active whenever we are exposed to a new visual scene or a new arrangement of distributed visual information in our environment. This assumption is in line with results from studies with simultanagnosia patients that are able to identify even complex objects but fail in the recognition of unfamiliar stimulus arrangements or alienated illustrations of regular objects (Dalrymple, Birmingham, Bischof, Barton, & Kingstone, 2010; Dalrymple, Bischof, Cameron, Barton, & Kingstone, 2009; Pavese, Coslett, Saffran, & Buxbaum, 2002; Robertson, Treisman, Friedman-Hill, & Grabowecky, 1997). Indeed, various aspects of higher vision underlie continuous learning mechanisms. Human neuroimaging studies demonstrated that response behavior and neuronal activity in regions associated with object perception—like the lateral occipital complex or fusiform face area—changed significantly for extensively trained object stimuli (Op de Beeck & Baker, 2010; Kourtzi & DiCarlo, 2006; Kourtzi, Betts, Sarkheil, & Welchman, 2005; Gauthier, Tarr, Anderson, Skudlarski, & Gore, 1999; Dolan et al., 1997). Other neuroimaging studies that investigated neuronal signal changes during learning of complex stimulus arrangements observed that perceptual training changes response characteristics in early visual (Zhang, Meeson, Welchman, & Kourtzi, 2010) and higher occipito-temporal/parietal regions (Mayhew, Li, & Kourtzi, 2012). These observations suggest complex neuronal dynamics underlying mechanisms of object and form perception.
The present experiment was designed to test the hypothesis that (bilateral) TPJ areas are involved in processing of novel stimulus arrangements requiring mechanisms of Gestalt perception, whereas complex but familiar stimuli are processed with less or no TPJ contribution. We conducted a learning experiment in which participants were repetitively exposed over 1 week to hierarchical stimuli in which a global Gestalt is perceived by the integration of local elements. Before and after the training period, the effects of stimulus characteristics were tested behaviorally and with fMRI. A measureable neuronal response to Gestalt perception training in area TPJ together with significant behavioral training effects would provide evidence for the specific role of this structure in the perception of novel stimulus configurations.
Twenty-four right-handed individuals (mean age = 26.1 years, SD = 2.7, 11 men) participated in this study. All had normal or corrected-to-normal vision, reported no history of neurological or psychiatric impairment, and gave their written informed consent. The study was approved by the local ethics committee.
Stimuli and Procedure
The whole investigation consisted of two fMRI sessions, two psychophysical test sessions, and 16 behavioral training sessions. The study started with the first (pretraining) fMRI measurement that was followed by a pretraining psychophysical test session outside the scanner on the next day. One week after the first fMRI measurement, a second (posttraining) fMRI measurement was performed, again followed by a psychophysical test on the next day. Between the pre- and postmeasurements, 16 training sessions were performed on 4 days (always four sessions/day).
Pre- and Posttraining fMRI Investigation
During the fMRI measurements and the psychophysical test sessions, participants were presented with three different kinds of stimuli (Figure 1). Two stimulus classes consisted of hierarchically organized stimuli where a global geometrical shape was built from local geometrical elements (Figure 1A, B). A third stimulus class consisted of black and white images of everyday objects (Figure 1C). From the three stimulus classes, we derived four experimental tasks: global perception of circles/squares (GCS; Figure 1A), global perception of triangles/stars (GTS; Figure 1B), local perception of circles/squares (LCS; Figure 1A), and object perception (OBJ; Figure 1C). From the two hierarchically organized stimulus classes, only one was used in the forthcoming training period (GCS; Figure 1A). The other hierarchically organized stimulus task (GTS; Figure 1B) as well as the local (LCS; Figure 1A) and object recognition task (OBJ; Figure 1C) served as controls.
The two sets of global/local stimuli were constructed as follows (see Figure 1): (A) global circles/squares constructed from local circles/squares or (B) triangles/stars that were created from small images of triangles/stars. Both sets consisted of four different combinations of local and global features (two congruent and two incongruent combinations). Each stimulus consisted of 900 small elements organized in 30 columns and 30 rows covering an area of 21.0° × 18.0° (width × height). The local elements had a size of 0.7° × 0.6°. To minimize spatial certainty and local learning effects, all global objects were presented at one of four different positions within an individual stimulus image (left top, right top, left bottom, right bottom; see Figure 1A, B). Furthermore, luminance and contrast were varied between the objects and their background (e.g., dark objects presented in light background and vice versa; see Figure 1A, B). To modulate global Gestalt perception, the global shapes were parametrically degraded by exchanging a proportion of 20%, 40%, 60%, or 80% of the small, local images across the respective global object images (GCS/LCS, GTS; see Figure 2). In correspondence to the procedure and findings by Huberle and Karnath (2012), the 20% scrambled condition represented “intact” perception of the global Gestalt and the 80% scrambled condition represented “disturbed” perception. For the object recognition task (OBJ; Figure 1C), we used 20 black and white images of everyday artificial (manmade) or natural objects derived from the Bank of Standardized Stimuli (Brodeur, Dionne-Dostie, Montreuil, & Lepage, 2010). Object stimuli were gradually superimposed with visual noise patterns to degrade perceptibility in correspondence with the scrambling of the global shapes (see Figure 2; OBJ). The average size of the depicted objects matched the average size of the global stimuli from the two sets of global/local stimuli.
All fMRI measurements consisted of six sessions with a duration of 9 min 16 sec each. In all four tasks (GCS, GTS, LCS, OBJ), participants had to perform a dichotomous decision. Via button presses on a single device with two buttons, they indicated whether they saw a global circle or square (GCS), a global triangle or star (GTS), a local circle or square (LCS), or an artificial or natural object (OBJ). In the GCS and GTS tasks, participants thus had to do a global perception task of a hierarchical form. In the LCS task, with the same stimuli as in GCS, local perception was required. In the OBJ task, they had to perform an object recognition task.
To integrate all four tasks in a feasible way, we used an event-related miniblock design (see Figure 3). For every fMRI session and every participant, the block sequence was identical. After an initial fixation period of 10 sec, the sequence of consecutive miniblocks began (see Figure 3). It started with the miniblock for GCS, followed by those for GTS, LCS, and OBJ. This series was repeated in the same order four times per fMRI session. Every miniblock started with a cue that was presented for 1500 msec and contained information about the following task. Additionally, it instructed the correct button responses for the presented stimuli, for example, left button press for a global square, right button press for a global circle in the GCS block shown in Figure 3. The left–right assignment of the responses for the respective stimuli and the hand used for the responses were kept constant throughout all behavioral and fMRI measurements for an individual participant and fully balanced across the participants. After a short fixation period of 1000 msec following the cue the actual task started. Every miniblock contained eight stimulus trials and two interleaved null events. Every experimental stimulus appeared for 300 msec, followed by a fixation period of 2700 msec. No stimuli were shown in the null events, which therefore consisted of a fixation period of 3000 msec. During the fixation period following stimulus presentation, participants were required to give a response by pressing one of the two buttons. We used interleaved null events to make the experiment less predictable for the participants, to provide a BOLD baseline measurement, and to jitter the time between successive stimuli and responses. With several limitations (no null events in direct succession, no null trials at the beginning or end of a miniblock, no events with same scrambling rate in direct succession), all four scrambling rates per task and the two blank periods were distributed in a pseudorandomized order in the respective miniblocks. Hence, every miniblock contained both target stimuli (e.g., global circles and squares) in all four scrambling rates. Furthermore, factors like congruency of global and local elements, global target stimulus position on the stimulus image, and luminance were distributed equally over all stimuli of one fMRI session. Every participant completed 128 trials of experimental stimuli within one fMRI session. Therefore, each fMRI measurement consisted of 768 trials.
To ensure that eye movement patterns did not differ between the four stimulus classes and the two fMRI measurements, we recorded eye movements during all fMRI sessions with an MR-compatible tracking device (MR-LR Sensomotoric Instruments, Teltow, Germany). Preprocessing of the eye tracking data included blink interpolation applying spline fitting algorithms, saccade detection, and smoothing of x and y positions. Afterwards, the absolute distance of gaze from the fixation dot was calculated for every sampled data point. These distances were sorted by task (GCS, GTS, LCS, OBJ). Gaze data of the whole miniblocks went into later data analysis; fixation periods (before and after the actual experiment) and cue events were discarded from the analysis.
Behavioral Testing and Learning Procedure
The same stimuli and tasks as in the scanner were used in the behavioral test measurements conducted on the day after the corresponding pre- and posttraining fMRI measurements. Stimuli were shown with the same size as in the scanner on a CRT monitor; the behavioral tasks were identical. Responses were collected with a standard keyboard where participants had to press arrow buttons for left or right. The left–right assignment for the respective stimuli was the same as in the scanner. The distance between the observant and the screen was kept constant with a chin rest. The four tasks (GCS, GTS, LCS, OBJ) were administered blockwise in four consecutive blocks of 12 min 32 sec. Each block comprised 288 experimental stimuli of one task. In this test, interleaved null events were also used to make the experiment less predictable. The number of null events was reduced by half, as no specific neuronal imaging parameters had to be taken into account. Similar design limitations as in the scanner (no null events in direct succession, no trials with same scrambling rate in direct succession) were applied to get a feasible pseudorandomized test design. Within every test block (e.g., GCS), factors like congruency, target position, and luminance were equally distributed over all stimuli that were presented in the respective block.
The learning sessions were conducted without a chin rest to provide more comfortable conditions for the participants. In total, 16 learning sessions that lasted for 16 min 20 sec were conducted. Four sessions were done consecutively within 1 day, resulting in four training days. The training days were randomly distributed over five possible days between the pretraining behavioral test and the posttraining fMRI measurement. In these learning sessions, participants were presented only with stimuli from the GCS task using an adaptive staircase scenario. The behavioral task in these sessions was the same as described above. Via button presses, participants had to indicate if they saw a global circle or square; key mapping was kept constant. In contrast to the test measurements, participants were provided with a feedback about their performance after each trial. Every learning session started with the easily perceivable 20% scrambling condition, and the difficulty of the task (i.e., scrambling rate of the stimuli) increased depending on the participant's performance. As soon as 10 consecutive trials reached a percent correct value of 70%, task difficulty was increased by 10% scrambling between scrambling levels of 20–50%. To measure behavioral improvements in perceptually demanding conditions more precisely, scrambling increased by only 2% as soon as a participant exceeded the 50% scrambling threshold. If participants performed in 10 consecutive trials worse than 30% correct, the task difficulty was reduced by 2%. Every training session lasted for a fixed number of 320 trials.
Four variables were analyzed as dependent variables for perceptual learning per training session: accuracy (ACC), RTs, maximum scrambling rate, and mean scrambling rate. “Maximum scrambling rate” is the maximum scrambling rate achieved in the respective training session; “mean scrambling rate” is the average scrambling rate from the same training session. ACC and RTs were averaged per participant and training session for scrambling rates that were achieved on the first training day. We only analyzed scrambling rates where the respective participant was able to perceive significantly above chance level (ACC ≥ 70%) at the end of the last training sessions of the first training day. For the analysis, trials of 20% scrambling were excluded because no behavioral learning effects for ACC were expected. For every participant, we thus had an individual profile of scrambling rates (e.g., nine scrambling levels from 40% to 62%) and corresponding ACC and RT values that were extracted from each training session. This resulted in a 16 (training sessions) × n (=number of scrambling levels that went into the analysis) matrix for ACC and RT. As ACC rates and RTs differed systematically regarding their absolute values between the different scrambling levels (lower ACC values, longer RTs for stimuli with higher scrambling rate), we normalized these variables. For every column of our ACC and RT matrices (representing a certain scrambling level), we subtracted the value of the first training session from the values of all 16 sessions. Therefore, these normalized values represent a comparable measure of learning for ACC and RT from the 16 training sessions over all scrambling rates. Finally, these normalized values were averaged for each participant over the respective columns (scrambling rates), resulting in two learning indices per training session representing learning effects for ACC and RT. In the end, we had four values quantifying behavioral learning (maximum and mean scrambling rate, ACC and RT indices) for each of our 16 training sessions for every participant. For an analysis of training days, we simply averaged the indices of the four training sessions conducted on the same day.
Functional MRI Data Acquisition and Analysis
We acquired EPI images with the following parameters: repetition time = 2000 msec; echo time = 35 msec; field of view = 192 × 192 mm; flip angle = 90°; 30 axial slices with a thickness of 3 mm, interleaved acquisition; matrix size = 64 × 64. In both fMRI measurements, a high-resolution T1-weighted anatomical image (1 × 1 × 1 mm3) was acquired from each participant. For all analyses of the fMRI data, we used the Statistical Parametric Mapping software package (SPM8; Wellcome Department of Imaging Neuroscience, London, UK; www.fil.ion.ucl.ac.uk/spm). At first, images recorded during pre- and postmeasurement fixation were discarded. Preprocessing of neuroimaging data involved spatial realignment to the mean image including unwarping. The mean EPI resulting from motion correction was coregistered to the anatomical image for every participant, and the respective transformations were applied to all functional images. The individual T1 anatomical images were segmented and normalized to the standard SPM T1 template. All EPI images were then normalized using parameters derived from the T1 unified segmentation and smoothed with a FWHM of 8 mm.
In the first-level analysis of each participant, we implemented a general linear model comprising 16 separate predictors for each experimental condition (20%, 40%, 60%, and 80% scrambling; four tasks) convolved with the hemodynamic response function as originally implemented in SPM8. Cue events were modeled as regressor of no interest, whereas fixation periods and null events were not modeled explicitly. This resulted in 23 regressors, including seven regressors of no interest comprising movement parameters from realignment and cue events. A high-pass filter with a cutoff period of 128 sec was applied to eliminate low-frequency noise components. A correction for temporal autocorrelation in the data was applied using an autoregressive AR (1) process.
We defined ROIs for analyses of neuronal effects of global perception learning. For a functional localization of global processing, we used results from two previous studies investigating neuronal correlates of Gestalt perception (Rennig et al., 2013; Huberle & Karnath, 2012). The functional ROIs from the study by Huberle and Karnath (2012) emerged from the same contrast with the same stimuli (GCS20% vs. GCS80%) and the same experimental task as in the present work and were successfully used for a functional reanalysis (Rennig et al., 2013; see Figure 4). Therefore, our ROIs derived from a previous study provide an independent localization of Gestalt perception in the human brain. The bilateral TPJ ROIs, originally from Huberle and Karnath (2012), come from a reanalysis of their data using SPM8 (Rennig et al., 2013). This reanalysis included the same steps as in the present fMRI analysis; this ensured maximum comparability between the two functional localizations of Gestalt perception. The bilateral TPJ ROIs (Rennig et al., 2013; Huberle & Karnath, 2012) were thresholded at p < .001 (uncorrected for multiple comparisons); further details about the reanalysis can be taken from Rennig et al. (2013). In the left hemisphere, two distinct functional ROIs, an anterior cluster and a posterior cluster, emerged from the reanalysis (Figure 4). According to structural labeling of the AAL atlas (Tzourio-Mazoyer et al., 2002), the anterior ROI overlapped anatomically with the supramarginal gyrus (SMG) and the superior temporal lobe (STL). The posterior ROI comprised mainly the angular gyrus, the middle temporal lobe, and the middle occipital gyrus and reached at its anterior borders marginally into the SMG. On the right hemisphere, a single cluster comprising aTPJ and pTPJ sections survived the statistical threshold of p < .001. This larger cluster comprised two local maxima that could be separated, comparable to the TPJ clusters in the left hemisphere, at a slightly higher threshold of p < .0008 (Figure 4). The two sections were easily separable according to the AAL atlas into an anterior section and a posterior section. Here, the anterior ROI comprised SMG and a small part of the STL, and the posterior ROI included the angular gyrus, middle temporal lobe, STL, and middle occipital gyrus.
In a previous study (Rennig et al., 2013), it was demonstrated that the left anterior TPJ area (aTPJ) responded stronger to global stimulus arrangements than an independent posterior TPJ section (pTPJ). Furthermore, evidence exists for an involvement of the right hemispheric aTPJ region, mainly comprising the SMG, in memory functions or target detection (Bzdok et al., 2013; Kubit & Jack, 2013). On the basis of reported functional differences between aTPJ and pTPJ sections (Bzdok et al., 2013; Kubit & Jack, 2013), clear anatomical allocations, the similarities between anterior and posterior clusters on the left and the fact that a slightly higher statistical threshold neatly separated two identifiable local maxima in the anterior part and posterior part of the right TPJ, we decided to analyze four ROIs that circumscribed the aTPJ and pTPJ in each hemisphere. MNI coordinates of the center of mass and size of the four ROIs were as follows: (R) aTPJ: x = 61.0, y = −38.8, z = 30.7; 2357 mm3; pTPJ: x = 45.8, y = −55.4, z = 25.6; 10605 mm3; (L) aTPJ: x = −58.0, y = −30.5, z = 26.5; 3060.0 mm3; pTPJ: x = −44.6, y = −58.8, z = 26.9; 9495 mm3 (see Figure 4). Once we specified conditions of interest, the ROI analysis was performed on the mean percent signal change (PSC), which was extracted using Marsbar SPM Toolbox from all voxels within the selected regions. We only used the PSC values from the “intact” (20% scrambled) and “disturbed” (80% scrambled) conditions from every stimulus task (GCS, GTS, LCS, OBJ) for the ROI analysis. With the PSC values, a 2 × 2 × 4 repeated-measures ANOVA with the factors (and levels) Measurement (pretraining vs. posttraining), Stimulus (intact vs. disturbed), and Task (GCS, GTS, LCS, OBJ) was performed for every ROI separately.
For a complementary exploration of the available data, we used individual contrast images obtained from the first-level analysis from each participant and each condition for a subsequent whole-brain analysis. Areas significantly involved in the perception of a global Gestalt were identified as those voxels showing significantly higher signals for 20% scrambled global shapes (“intact” global perception) compared to 80% scrambled shapes (“disturbed” global perception). Therefore, for the whole-brain analysis, we calculated three different contrasts from the pretraining fMRI measurement: GCS20% vs. GCS80%, GTS20% vs. GTS80%, GCS20% + GTS20% vs. GCS80% + GTS80%.
Behavioral Testing and Learning Procedure
The behavioral test results outside the scanner were consistent with those collected in the scanner. We thus present the behavioral data from the fMRI sessions, as these are more relevant for the interpretation of our neuroimaging results. We calculated separate 2 × 4 × 4 repeated-measures ANOVAs with the factors Measurement (pre- vs. posttraining), Task (GCS, GTS, LCS, and OBJ), and Stimulus (20%, 40%, 60%, 80%) for RT and ACC values. We observed a significant three-way interaction of all factors for ACC (F(9, 15) = 4.59, p = .005) and RT (F(9, 15) = 2.94, p = .032). All other main effects and two-way interactions were also significant (p < .05). We thus tested each experimental factor level of our 4 × 4 (Task, Scrambling) design pretraining against posttraining. Thus, we performed 16 t tests per dependent variable that were corrected for multiple comparisons, applying Bonferroni correction. For ACC, three comparisons from the GCS (20%, 40%, 60% scrambling; T(23) = 4.11, p < .001; T(23) = 4.27, p < .001, T(23) = 5.92, p < .001), one from the GTS (60%; T(23) = 3.53, p = .002), and two from the OBJ task (20%, 60% scrambling; T(23) = 3.95, p = .001; T(23) = 6.11, p < .001) showed significant differences between pre- and posttraining measurements. All other tests did not reach significance even without Bonferroni correction. For RT, all comparisons but those for GTS 80% scrambling and all LCS levels showed significant results. However, without a Bonferroni correction, all individual comparisons showed significant results (p < .05). As RTs decreased for virtually all tasks, only the behavioral data for ACC are illustrated in Figure 5. The results illustrate a perceptual improvement that was most prominent in the trained global perception task.
The results during the behavioral training are illustrated in Figure 6. They indicate strong behavioral improvements over several variables measuring perceptual abilities of global Gestalt processing. We analyzed maximum scrambling rate, mean scrambling rate, ACC, and RT for training sessions and days. To analyze learning effects over sessions, we calculated linear regressions for every participant and each of the four dependent variables over the 16 training sessions. We used individual beta and R2 values to calculate one-sample t tests to demonstrate significant deviations of the regression line from zero. For all four variables, these tests showed significant results for beta and R2 (maximum scrambling rate: beta: mean = .24, T(23) = 5.05, p < .001, R2: mean = .21, T(23) = 5.31, p < .001; mean scrambling rate: beta: mean = .20, T(23) = 5.63, p < .001, R2: mean = .27, T(23) = 6.04, p < .001; ACC: beta: mean = .004, T(23) = 5.31, p < .001, R2: mean = .23, T(23) = 5.37, p < .001; RT: beta: mean = 2.66, T(23) = 3.61, p = .002, R2: mean = .25, T(23) = 5.38, p < .001). To further analyze the general effects across training days, neglecting within-day variability, we averaged the results of the four training sessions held on one day and conducted the same analysis over training days as we did over sessions. Also in this analysis, one-sample t tests for the four variables showed significant results for beta and R2 (maximum scrambling rate: beta: mean = .95, T(23) = 5.22, p < .001, R2: mean = .49, T(23) = 7.86, p < .001; mean scrambling rate: beta: mean = .77, T(23) = 5.68, p < .001, R2: mean = .56, T(23) = 8.80, p < .001; ACC: beta: mean = .014, T(23) = 4.95, p < .001, R2: mean = .52, T(23) = 7.21, p < .001; RT: beta: mean = 8.97, T(23) = 3.31, p = .003, R2: mean = .46, T(23) = 6.32, p < .001).
We further tested our behavioral data in terms of “stimulus type” (e.g., global circle vs. square) and “congruency” looking for possible effects evoked by the nature of the applied task or stimulus construction. These analyses showed no effects possibly biasing our main analyses of behavioral or neuronal data. Furthermore, these analyses demonstrated that our hierarchical stimuli (GCS, GTS, LCS) had the expected characteristics of typical global/local stimuli and evoked the global precedence effect (Navon, 1977). The analyses and results can be inspected in the supplementary results section.
During all stimulation periods, participants were able to fixate properly and did not exceed the central fixation area (±3° visual angle in x and y direction) during stimulus presentation. To inspect the data for systematic differences between the tasks or measurements, we calculated the mean distance between gaze position and the fixation cross across all miniblocks for each task, separately for the two measurements and every participant. With this variable, we calculated a 2 × 4 repeated-measures ANOVA with the following factors and levels: Measurement (pre- vs. posttraining) and Task (GCS, GTS, LCS, and OBJ). This ANOVA showed no significant interaction (F(3, 20) = 0.70, p = .56) or main effects (Measurement: F(1, 22) = 0.04, p = .38; Task: F(3, 20) = 0.35, p = .79).
We performed ROI analyses with mean PSC extracted from our four TPJ regions associated with global processing (see Figure 4; see section Functional MRI Data Acquisition and Analysis). For each ROI, we performed a 2 × 2 × 4 repeated-measures ANOVA with the following factors and levels: Measurement (pretraining vs. posttraining), Stimulus (intact vs. disturbed), and Task (GCS, GTS, LCS, OBJ). For the right hemispheric aTPJ ROI, we observed a significant three-way interaction effect of Measurement, Stimulus, and Task (F(3, 21) = 3.51, p = .033), a significant two-way interaction of Measurement and Stimulus (F(3, 21) = 10.87, p = .003), and a significant main effect for Task (F(1, 23) = 18.21, p < .001). On the basis of the significant three-way interaction, we performed four two-way ANOVAs separately for every task with the factors Measurement (pretraining vs. posttraining) and Stimulus (intact vs. disturbed). For the trained task (GCS), we observed a significant interaction of Measurement and Stimulus (F(3, 21) = 14.33, p = .001). For the GTS task, a nearly significant interaction of Measurement and Stimulus (F(3, 21) = 3.62, p = .070) was evident. For LCS and OBJ, no main effects or interactions came close to a significant result (p > .10). Even with a Bonferroni correction for all four ANOVAs resulting in a p threshold of .0125, the interaction for the trained GCS task can still be considered as significant. The PSC results of the four tasks and the two measurements are illustrated in Figure 7. On the basis of the significant two-way interaction for GCS, we performed four paired t tests comparing PSC values for “intact” and “disturbed” (20% and 80% scrambling) and “pretraining” and “posttraining” measurements. At first, we compared the two scrambling rates per measurement. We observed a significant difference comparing PSC values for “intact” and “disturbed” (20% vs. 80% scrambling) and for the “pretraining” (T(23) = 4.89, p < .001) but not for the “posttraining” measurement (T(23) = −1.50, p = .15; see Figure 7, GCS). Comparing PSC values for the two scrambling rates between the two measurements revealed a significant result for “intact” (20% scrambling; T(23) = 2.96, p = .007) and a marginally significant result for “disturbed” global perception (80% scrambling; T(23) = −2.13, p = .044; see Figure 7). These results clearly illustrate learning-dependent changes that were specific for the trained global perception task (GCS). We demonstrate a significant decrease for “intact” and a significant increase for “disturbed” global perception task (see Figure 7). Furthermore, the significant difference in the aTPJ in the pretraining measurement indicating a significant involvement of this structure in Gestalt perception disappeared (verified by two- and three-way interactions; see also Figure 7). In the right hemispheric pTPJ ROI, we observed significant main effects for Task (F(3, 21) = 5.63, p = .005) and Stimulus (F(1, 23) = 13.15, p = .001). The 2 × 2 × 4 ANOVA for the left hemispheric aTPJ ROI revealed no significant results. In the left pTPJ ROI, we observed significant main effects for Measurement (F(1, 23) = 9.83, p = .005) and Stimulus (F(1, 23) = 19.51, p < .001). These results indicate that specific learning dependent changes can exclusively be attributed to the right hemispheric aTPJ ROI, whereas the other ROIs did not respond significantly to perceptual training. The four TPJ ROIs (a/pTPJ in the left/right hemisphere) differed substantially regarding their size (min = 2357, max = 10,605 mm3; see above and Figure 4). To ensure that the results were not simply caused by simple size differences of our ROIs, we performed an ROI analysis with spherical ROIs (radius = 8 mm) constructed around the respective peak voxels of the clusters. These analyses revealed comparable results and showed that the original results did not depend on mere ROI size. A detailed overview can be inspected in the supplementary results section.
To ensure that the results are specific for our distinct ROI, we performed the same analysis in four bilateral brain regions associated with visual attention or expertise. The analyses and results can be inspected in the supplementary results section.
With the data from the pretraining fMRI measurement, we performed three contrasts to compare “intact” to “disturbed” global perception before training: GCS20% vs. GCS80%, GTS20% vs. GTS80%, GCS20% + GTS20% vs. GCS80% + GTS80%. Two of these contrasts (GCS20% vs. GCS80%, GCS20% + GTS20% vs. GCS80% + GTS80%) clearly revealed posterior temporoparietal brain areas as crucial regions of Gestalt perception (Figure 8). The comparison of “intact” versus “disturbed” Gestalt perception over both global perception tasks (GCS, GTS) showed a significant involvement of the right TPJ for “intact” global perception (p < .05, FWE-corrected; Figure 8A). The same comparison with a more liberal statistical threshold (p < .001, uncorrected; Figure 8A) revealed bilateral TPJ activations as well as activity in bilateral cingulate regions and right hemispheric inferior frontal cortex, superior parietal lobe, and BG (not visible in Figure 8A). A comparison of “intact” versus “disturbed” perception restricted to the GCS task revealed bilateral TPJ areas as the neuronal correlate of Gestalt perception only for an uncorrected statistical threshold of p < .001 (Figure 8B). Applying the same contrast separately for the GTS task, the same bilateral TPJ regions (besides several other activation clusters) were observable only for a very liberal statistical threshold (p < .01; results not shown). The results are in good agreement with evidence from previous studies showing a significant involvement of (bilateral) TPJ areas in global Gestalt processing (Rennig et al., 2013; Huberle & Karnath, 2012; Himmelbach et al., 2009). A detailed overview of the results can be inspected in the supplemental results section.
This study investigated the role of bilateral TPJ regions in global Gestalt perception. On the basis of the previously reported functional differences between aTPJ and pTPJ sections (Bzdok et al., 2013; Kubit & Jack, 2013; Rennig et al., 2013), we asked if aTPJ areas are mainly involved in the processing of novel complex stimuli. A behavioral training over 1 week familiarized participants with complex global stimulus material. We hypothesized that increasing familiarity with the test stimuli would change response characteristics of aTPJ areas pre- and posttraining. On the behavioral level, we observed clear improvements in the trained global perception task (GCS), whereas for the untrained global perception task (GTS), only slight training effects were evident. Over the 16 training sessions and four training days, participants showed a continuous decrease of RTs and increasing ACC values, indicating enhanced ability to integrate global visual arrangements. No significant behavioral changes were observed for the untrained local perception task (LCS), whereas in the untrained object perception task (OBJ) moderate behavioral changes were evident. It is possible that the slight improvement in the latter condition is due to simple memory effects evoked by repeated standardized testing with identical stimuli from a limited stimulus set. The behavioral results are in good agreement with studies showing a partial specificity of learning for trained (object) stimuli (Baeck & Op de Beeck, 2010; Furmanski & Engel, 2000; Grill-Spector, Kushnir, Hendler, & Malach, 2000; Sigman & Gilbert, 2000). In these studies, participants showed clear training effects on the trained stimulus class whereas learning effects were less pronounced for untrained but similar stimuli.
The comparison of pre- and posttraining BOLD signals in the delineated ROIs demonstrated for the first time training effects in area TPJ. Significant changes were observed in the right hemispheric aTPJ ROI for the trained global perception tasks (GCS). In the control tasks requiring untrained global perception (GTS), local processing (LCS), and object recognition (OBJ), no systematic signal modulations were observed. In the right hemispheric pTPJ and both left hemispheric TPJ ROIs, no statistically significant effects were evident. In conclusion, the signal changes argue for an involvement of the right hemispheric aTPJ region in processing of mainly novel complex stimulus configurations. With increasing familiarity for the tested stimuli, this TPJ section showed fundamentally different response characteristics. The results are in good agreement with observations in patients with simultanagnosia. Although even complex familiar objects can be recognized, these patients fail in the identification of novel stimulus arrangements or alienated (unfamiliar) illustrations of regular objects (Dalrymple, Birmingham, et al., 2010; Dalrymple et al., 2009; Pavese et al., 2002; Robertson et al., 1997). The present findings suggest that this observation occurs due to learning dependent signal modulations.
Our present observations fit well with the previous findings by Rennig et al. (2013), where it was demonstrated that the right aTPJ area responded stronger to global stimulus arrangements than an independent pTPJ section. In this case, where TPJ responses of visual experts were compared to those of novices, the aTPJ seemed to be more sensitive to extensive exposure of visual material supporting Gestalt-like perception of complex visual arrangements. Moreover, in the pretraining fMRI measurement, we were able to replicate the results from the study of Huberle and Karnath (2012) using identical (GCS) or similar (GTS) stimulus material but applying a fundamentally different fMRI procedure (event-related miniblock design; see Methods section; Figure 8). This further strengthens the assumption that area TPJ represents a crucial region for global Gestalt processing (Rennig et al., 2013; Huberle & Karnath, 2012; Himmelbach et al., 2009). Although our whole-brain analysis for the GCS task revealed bilateral TPJ regions only for an uncorrected statistical threshold, the results can be considered as very consistent. This study design contained more experimental conditions resulting in a reduced statistical power. Anyway, based on a strong hypothesis from previous studies (Rennig et al., 2013; Huberle & Karnath, 2012; Himmelbach et al., 2009), the present results argue for high validity of the approach and the results.
We assume that the observed neuronal signal changes in area TPJ correspond to a more efficient processing of “intact” global stimuli and a higher sensitivity for degraded but potential global targets (Kourtzi et al., 2005). In any case, the training-induced enhancements of neuronal responses for “disturbed” global perception observed in this study are in line with neuroimaging studies indicating that visual learning of degraded (George et al., 1999; Dolan et al., 1997), masked (Grill-Spector et al., 2000; James, Humphrey, Gati, Menon, & Goodale, 2000) or noise-embedded (Kourtzi et al., 2005; Schwartz, Maquet, & Frith, 2002) targets increases neuronal signals. Likewise, neurophysiological studies have suggested that training with low-salience targets or objects in cluttered scenes leads to stronger neuronal signals indicating a higher sensitivity to target features and facilitation for the detection and integration of a (potential) global form (Rainer, Lee, & Logothetis, 2004; Kobatake, Wang, & Tanaka, 1998; Tovee, Rolls, & Ramachandran, 1996; Logothetis, Pauls, & Poggio, 1995; Sakai & Miyashita, 1991). This enhanced neuronal sensitivity can be explained as an increased internal signal-to-noise ratio for trained stimuli supporting the selection of a global shape (Dosher & Lu, 2006). In contrast, lower neuronal responses observable for “intact” global processing after training indicates more efficient neuronal processing for high-salience, unambiguous targets. This effect is known from previous neuroimaging studies investigating perceptual learning effects on pop-out targets (Kourtzi et al., 2005; Chao, Weisberg, & Martin, 2002; Koutstaal et al., 2001; Henson, Shallice, & Dolan, 2000; Jiang, Haxby, Martin, Ungerleider, & Parasuraman, 2000; Van Turennout, Ellmore, & Martin, 2000). These effects were confirmed by similar results from neurophysiological studies in monkeys (Schoups, Vogels, Qian, & Orban, 2001). Especially, a study by Kourtzi et al. (2005), which investigated perceptual learning with shapes arranged from Gabor elements showed interactions between stimulus saliency and learning-induced neuronal activation changes. It was demonstrated that trained shapes that were difficult to perceive because of a fuzzy background produced higher neuronal responses than untrained versions of these stimuli in early and higher visual areas. In contrast, trained shapes that were easy to perceive showed a lower neuronal signal compared to untrained ones in higher visual areas. A recent study investigating TPJ involvement in the perception of gratings (Beauchamp, Sun, Baum, Tolias, & Yoshor, 2012) confirmed these observations as well as our present results. It was demonstrated that electrical stimulation of human TPJ areas enhanced detection rates for low-salience stimuli whereas perception in undisturbed viewing conditions was unaffected. In general, our results are in good agreement with existing evidence on neuronal effects of visual learning. Furthermore, we do not attribute the observed signal modulations in area TPJ to mere changes in visual attention, because no systematic activation changes were observable in bilateral FEFs. Moreover, the nature of our stimuli that varied in contrast, position, and coloring suggests that local processing of single elements or object parts was not trained but actually visual top–down processing in the sense of Gestalt perception.
Regarding the general role of the TPJ in Gestalt perception, Fink et al. (1996, 1997) demonstrated that bilateral TPJ regions are crucial areas for attentional shifts between global and local aspects of Navon letters. A study by Yamaguchi and colleagues (2000) supported this view. Here, the right hemispheric TPJ showed significant activation during the cuing phase for global cues but not for local ones. A different function was addressed to the right hemispheric TPJ area by a study of Weissman and Woldorff (2005): TPJ was identified as a region responsible for maintaining a global percept whereas attentional control about local or global perception in Navon letters was controlled by the intraparietal sulcus. Moreover, this specific function of the intraparietal sulcus is supported by other neuroimaging (Weissman, Mangun, & Woldorff, 2002) or TMS studies (Romei, Driver, Schyns, & Thut, 2011). The results about a crucial involvement of TPJ areas for mere perception of global structures by Weissman and Woldorff (2005) are in line with further studies investigating neuronal mechanisms of global processing (Rennig et al., 2013; Huberle & Karnath, 2012; Himmelbach et al., 2009). All these studies revealed a significant involvement of bilateral TPJ areas in global Gestalt perception whereas no attentional shift from local to global or vice versa had to be performed by the participants. These findings are in line with results of the current study and emphasize an involvement of bilateral TPJ regions in mere perceptual mechanisms of Gestalt processing.
However, alternative explanations such as a shift of neuronal processing for extensively trained global stimuli from visual integration in area TPJ to other regions—of course—are also plausible. For example, a previous study investigating neuronal training effects for a visual search task demonstrated training dependent parietal and lateral occipital signal decreases in favor of an increase in early visual areas (Sigman et al., 2005). This activation change was interpreted as a redistribution of the functionality of different cortical areas involved in object identification. In this study, a possible shift in neuronal activation may have occurred from integration-related processes in area TPJ in favor of a stronger ventral involvement and mechanisms of object processing.
We conclude that (anterior right hemispheric) TPJ regions are involved in processing of mainly novel global stimuli. For the first time, we showed that fMRI signals in TPJ regions are modulated through extensive perceptual training with complex global configurations. With increasing familiarity, these areas changed sensitivity and selectivity for complex stimulus arrangements. The findings thus strengthen the view about the (right hemispheric) TPJ as a crucial module for Gestalt perception (Rennig et al., 2013; Huberle & Karnath, 2012; Himmelbach et al., 2009).
This work was supported by the DFG (Ka 1258/10-1) and the European Union (ERC StG 211078). We thank Ida Zündorf for her assistance during fMRI data acquisition.
Reprint requests should be sent to Johannes Rennig, Center of Neurology, Division of Neuropsychology, University of Tübingen, D-72076 Tübingen, Germany, or via e-mail: firstname.lastname@example.org.