Children Use Regions in the Visual Processing and Executive Function Networks during a Subsequent Memory Reading Task

Abstract

Memory encoding is a critical process for memory function, which is foundational for cognitive functioning including reading, and has been extensively studied using subsequent memory tasks. Research in adults using such tasks indicates the participation of visual and cognitive-control systems in remembered versus forgotten words. However, given the known developmental trajectories of these systems, the functional neuroanatomy of memory encoding in children may be different than in adults. We examined brain activation for silent word reading and checkerboard viewing during an event-related reading task in 8–12 year-old children. Results indicate greater activation for checkerboard viewing than lexical processing in early visual regions, as well as for lexical processing versus checkerboard viewing in regions in left sensorimotor mouth, cingulo-opercular and dorsal-attention networks. Greater activation for remembered than forgotten words was observed in bilateral visual system and left lateralized regions within the ventral- and dorsal-attention, cingulo-opercular and fronto-parietal networks. These findings suggest a relatively mature reliance on the cognitive-control system, but greater reliance on the visual system in children when viewing words subsequently remembered. The location of regions with greater activity for remembered words reinforces the involvement of the attention and cognitive-control systems in subsequent memory in reading.

Memory encoding is defined as the process by which an experience is consolidated for later retrieval. Determining the neural systems involved in encoding written information was the focus of the current study.

The first studies examining the effect of subsequent memory, i.e. the differential neurobiological signature for items remembered vs forgotten, showed that differential neural processing at the time of encoding determined how well information was remembered (Paller et al. 1987; Kutas 1988; Brehmer et al. 2016; Leventon and Bauer 2016). These studies examined the neurobiological signatures for remembered vs forgotten items [emotional (Leventon and Bauer 2016), verbal (Campeanua et al. 2014), written (Paller et al. 1987; Kutas 1988; Wagner et al. 1998; Brehmer et al. 2016)] and had consistent findings showing greater event related potential (ERP) or activation (as indicated by fMRI) for the remembered items. Those studies were focused on subsequent memory during reading and revealed greater amplitude components related to early processing stages for remembered words than forgotten words, suggesting that the processing of the to-be-remembered words differs from the forgotten ones (Paller 1990). These studies also revealed greater amplitude in left lateralized components for remembered words (Paller et al. 1987), especially in the left frontal cortex, when both adult (Wagner et al. 1998) and child participants are training on the presented written words (Brehmer et al. 2016). This highlights the greater involvement of left hemisphere regions in a subsequent memory task (Paller et al. 1987; Brehmer et al. 2016). However, the neural networks involved in subsequent memory during word reading in children are yet to be defined.

Reading, the cognitive ability that allows humans to translate written graphemes to their corresponding phonemes in a fluent and accurate manner, is not as intuitive as one might think. Synchronization between several modalities—phonological processing, as well as visual abilities and cognitive control—is essential for a successful reading process (Pugh et al. 2000; Shaywitz and Shaywitz 2003, 2008; Horowitz-Kraus and Hutton 2015). The Parallel Distributed Processing model for reading acquisition supports this complexity by suggesting a parallel involvement of phonological, orthographical, and semantic processors during reading development (Seidenberg and McClelland 1989). Young children are exposed to spoken language and narratives that are foundational for future phonological processing, which is essential for reading (Berl et al. 2010; Horowitz-Kraus et al. 2013). Orthographic processing is involved in recognizing first the letters and then the words, eventually in a holistic effortless manner, and then matching the appropriate semantic meaning (Noble and McCandliss 2005); for review of neuroimaging studies in reading see also (Fiez and Petersen 1998). Recent studies have indicated the involvement of cognitive control in these processes (Horowitz-Kraus et al. 2017; Patael et al. 2018). This includes working memory, visual attention, and speed of processing, all of which are critical components allowing for the fast and automatic processing required in reading; for review see (Horowitz-Kraus et al. 2017). In one of the first neuroimaging studies describing the brain’s reading systems, Pugh et al. (2000) suggested that the angular gyrus, fusiform gyrus, and inferior frontal gyrus are key regions involved in reading. More recent reports cite the cerebellum as being important to reading (Vlachos et al. 2007; Travis et al. 2015). Further neuroimaging work has confirmed that all of these regions are part of neural systems supporting the reading process, such as those related to language processing (Horowitz-Kraus et al. 2013), visual processing (Vogel et al. 2012), and cognitive-control systems; specifically ventral attention, visual attention, cingulo-opercular and fronto-parietal regions (Schlaggar and McCandliss 2007; Horowitz-Kraus, Toro-Serey et al. 2015). These networks were found to be involved in improved reading ability both in typical readers and in those with dyslexia, whereas greater functional connectivity within these networks was related to improved reading ability (Horowitz-Kraus, DiFrancesco et al. 2015; Horowitz-Kraus, Toro-Serey et al. 2015). Others reported the involvement of parts of the fronto-parietal network during improved reading comprehension in children with reading difficulties (Roe et al. 2018). The authors described the role of executive functions and attention abilities in intact reading; a question arises as to the involvement of regions within executive function and attention networks as part of the underlying factors contributing to word consolidation. This was examined in the current study.

The ultimate goal of reading is to comprehend written material. The Samuels and LaBerge model for reading suggested that only when reading becomes automatic (i.e., phonological and orthographical processes are effortless) are cognitive-control resources then available for comprehension (LaBerge and Samuels 1974). Neuroimaging studies also have explored the neural systems involved in words read that were then encoded into memory (i.e., remembered) vs words forgotten in adult typical readers (Wagner et al. 1998; Mei et al. 2010). The Wagner study used an event-related design to show that remembered words generated greater activation than forgotten words in left prefrontal cortex regions (left frontal operculum, left anterior, and posterior inferior frontal gyrus) and left temporal regions (left parahippocampal gyrus and fusiform gyrus). It also has been suggested that frontal systems participate in organizing written information in working memory prior to processing it in the hippocampus for memory consolidation (Vogel et al. 2012). Additional studies demonstrated greater activation of regions in the visual system, such as the left fusiform gyrus, for remembered words and faces in adults (Mei et al. 2010), as well as verbal vs pictorial information remembered in a meta-analysis of 74 studies (Kim 2011). However, since cognitive-control abilities improve with development (Giedd 2004) and the sensitivity of the visual system to written material improves with reading experience (Olulade et al. 2013), a question arises as to the involvement of cognitive control and visual systems for remembering words in children.

Exploring the neural systems involved in successful reading will facilitate an understanding of both superior as well as impaired reading. The current study focused on the neural systems involved in words remembered vs forgotten in children who are typical readers. We used an intentional encoding task, i.e., a reading task that included a memory manipulation to encourage attentiveness to the words, with warning of a subsequent memory test. Based on the findings observed in adults (Mei et al. 2010), we hypothesized that children would engage the visual system to a greater extent for words they remembered than those they forgot. We also anticipated that children would use the cognitive-control system more for remembered words than those forgotten. This includes networks related to alerting and orienting attention to the word stimuli (e.g., ventral and dorsal attention) and those related to top-down internal processing (i.e., fronto-parietal and cingulo-opercular networks); see (Petersen and Posner 2012). We used the Power study network boundaries to categorize our fMRI results (Power et al. 2011). These networks were defined using resting-state functional connectivity boundary maps that represent cortical regions (Power et al. 2011) also found during event-related fMRI tasks (Shine et al. 2016).

Methods

Participants

A total of 33 typically reading children 8–12 years of age consented to participate in the study. All participants were right-handed Caucasian, native English speakers with average socioeconomic status (as reported by the families). They displayed normal or corrected-to normal vision in both eyes, and had normal hearing. None had a history of neurological or emotional disorders. Parents reported on the absence of attention difficulties, which was verified by the Sky Search test from the Test of Everyday Attention for Children (TEA-Ch) battery of tests (Manly et al. 1999). Participants were recruited through posted ads and commercial advertisements. All participants signed informed written assent and their parents provided informed written consent prior to inclusion in the study, and all were compensated for participation. The Cincinnati Children’s Hospital Medical Center Institutional Review Board approved the study. After removing data for the 2 participants who did not complete the task reported in this study, data for 31 participants were analyzed. Three of these sets of data had several timecourses that did not have the shape of a hemodynamic response function, and one was insufficient after motion scrubbing. Therefore, the results of this study are based on the data for 27 participants (14 males, 13 females).

Behavioral Measures

All study participants were assessed for nonverbal, verbal, and attention abilities using the Test of Nonverbal Intelligence (TONI) (Brown et al. 1997), as well as for verbal abilities using the Peabody picture vocabulary task (Dunn and Dunn 2007) and the TEA-Ch Sky Search test (Manly et al. 1999).

Reading Measures

Reading was evaluated using a battery of normative reading tests in English. Inclusion criteria was a score of 50th percentile and higher in all reading tasks. The reading battery included: (1) Automatic word reading accuracy/orthography: Test of Sight word reading efficiency (TOWRE-SWE) (Torgesen et al. 1999), (2) Automatic decoding: Pseudo-word reading efficiency subtest (TOWRE-PWE) (Torgesen et al. 1999), (3) Non-timed word reading accuracy/orthography (letter-word subtest) (Woodcock and Johnson 1989), (4) Non-timed decoding of pseudo-word reading (word-attack subtest) (Woodcock and Johnson 1989), (5) Phonemic awareness: the Comprehensive test of phonological processing (CTOPP, elision subtest) (Wagner et al. 1999), (6) Reading fluency: the test of silent reading efficiency and comprehension (TOSREC) (Wagner et al. 2010), and (7) Reading comprehension (Woodcock and Johnson 1989). Averages and standard deviations for each test were generated.

Neuroimaging Data

Functional MRI event-related word-reading task

All children performed a 9-minute event-related functional MRI scan as part of a longer scanning session. Thirty-nine high-frequency encountered words were presented for 1700 msec each, interspersed with 39 stationary checkerboards presented for 1700 msec each. A fixation cross appeared when neither stimulus was present for a total of 263 two-second image volumes. Two of the words were presented twice, for a total of 41 word presentations. Participants were asked to read the words silently and to try to remember them. The children were also told that after the scan they would have to identify the words they read in the scanner. Outside the scanner, participants were presented with 37 of the words from the scanner task and an additional 22 new words matched for frequency and length. From the randomized listing of the 59 words, participants were asked to circle the words they had read in the scanner task. %hits was calculated from the number of correctly circled words. Accuracy d’ was calculated from normalized proportion of hits minus normalized proportion of false alarms. To determine the relationship between the in-scanner d’ accuracy and %hits measures and tested reading measures, a Pearson correlation was conducted.

MRI acquisition and data preprocessing

All of the participants were scanned on a 3T Phillips Achieva MRI scanner using a 32-channel head-coil. A gradient echo planar sequence was used for T2*-weighted functional MRI scans with the following parameters: TR/TE = 2000/38 ms; BW = 125 kHz; FOV = 25.6 × 25.6 cm; matrix = 64 × 64; slice thickness = 5 mm, 40 slices. 263 whole-brain volume scans were taken during the word-reading event-related task for a total imaging time of 8.8 min. A high-resolution T1-weighted 3D anatomical scan was acquired using an inversion recovery (IR)-prepared turbo gradient-echo acquisition protocol with a spatial resolution of 1 × 1 × 1 mm³. To insure comfort during the scan, participants were desensitized to the scanner condition (Byars et al. 2002). Head motions were controlled using elastic straps that were attached to either side of the head-coil apparatus.

Pre and post processing of data

Functional images were processed to remove noise and artifacts using a series of automated steps (Miezin et al. 2000). Slice by slice normalization corrected for differences in image intensity caused by interleaved acquisition, and sync interpolation corrected for the different times of acquisition. Rigid body translation and rotation realigned all images to the first collected (Snyder 1996). The mode voxel value was normalized to 1000 to allow for across-scan comparison. Each participant’s T2* images were transformed into Talairach atlas space (Talairach and Tournoux 1988) using the participant’s T1 image as an intermediary to allow direct statistical comparison.

Blood oxygen level dependent (BOLD) activity related to the trials was modeled using a general linear model (GLM), estimating a value for the 8 MR frames (16 seconds) following the stimulus presentation rather than assuming a hemodynamic response shape (Friston et al. 1994; Josephs et al. 1997; Miezin et al. 2000; Ollinger et al. 2001). Two separate GLMs were computed for each participant—one that used words and checkerboard presentations as the only 2 events and another that estimated remembered and forgotten words separately, as well as checkerboard presentations. All GLMs also estimated a baseline and trend term. To help remove motion artifact, frames of data were removed from the GLM estimation if the frame-to-frame displacement was greater than 0.9 mm (Siegel et al. 2014). These motion-scrubbed GLMs were used in the final region-wise analysis of variance (ANOVA).

ANOVA was performed on each set of GLMs, all in Talairach atlas space with a voxel size of 2 mm on a side. The first compared the timecourses for word vs checkerboard presentations, resulting in, among other effects, a Main Effect of Time (MET) image showing which brain voxels’ hemodynamic response differed from baseline when combining both presentation types, and a Task × Time image showing differences between words and checkerboards. In a second ANOVA, we compared remembered and forgotten words. Two images showed (1) brain voxels that were active for all word presentations (Memory MET or MMET) and (2) which voxels differed between words that were remembered and forgotten (Memory × Time).

To determine what effects drove significance in the ANOVA images, regions were chosen from the 4 described images using an in-house algorithm (https://readthedocs.org/projects/4dfp/). This automatic algorithm smoothed the images with a 4 mm Gaussian kernal, dropped 10 mm spheres on each peak with a z-value greater than 3, combined peaks closer than 10 mm apart into a single region, then removed any voxels not in the Monte Carlo-based multiple comparisons corrected image (required 45 contiguous voxels with z > 3).

Region-wise ANOVAs then computed participant timecourses in each region for 4 conditions: (1) all words, (2) checkerboards, (3) remembered words, (4) forgotten words. These ANOVAs also computed statistical comparisons on the timecourses, asking: (1) Do the combined stimuli cause a difference from baseline? (MET, words and checkers); (2) Is there a difference between words and checkerboards? (Task × Time, words vs checkers); (3) Do words alone cause a difference from baseline? (MMET, all words); (4) Is there a difference between words that were remembered and those that were forgotten? (Memory × Time, remembered vs forgotten). Here, we report regions that had a statistically significant Memory × Time interaction effect and a biologically plausible timecourse. Many overlapping regions resulted from the different images. For the sake of clarity, we chose only one of these overlapping regions; in each case, we chose a region from one of the main effect of time images, either words-and-checkers (MET) or all-words (MMET), to avoid the accusation of “double dipping” by choosing regions from the Memory × Time image. We also removed regions with timecourses that did not have the shape of a hemodynamic response function, as well as those with large amplitudes presumed to be in draining veins.

For the functional context, each of these regions, as well as the Power reported network boundaries (Power et al. 2011) were placed on an inflated cerebral surface using Workbench (Marcus et al. 2011) (https://www.humanconnectome.org/software/connectome-workbench).

To relate the accuracy rate of words recognized in the post-scan test to reading level, Pearson correlations compared the hits percentage (# words recognized/number of words presented in the scanner) and d’ (normalized proportion of hits minus normalized proportion of false alarms) in the post-scan memory test with the standardized reading tests outlined in the Behavioral Reading Measures section.

Results

Behavioral Results

Participants demonstrated intact nonverbal ability, as well as reading and reading comprehension abilities in all examined reading measures (Table 1).

Table 1.

Mean and standard deviations of reading measures (N = 27; n = 14 males, n = 13 females)

	Mean	Standard deviation	Skewness
Age	10.03	1.18	0.22
General non-verbal ability; TONI (Brown et al. 1997), percentile	56.79	19.54	0.353
Verbal ability; PPVT (Dunn and Dunn 2007), percentile	75.0	18.53	−0.659
Attention ability; Conners Parent (Conners 1989), T Score	50.11	11.08	2.422
MRI accuracy rate; percentage	63.27	17.28	−0.347
Reading fluency; TOSREC (Wagner et al. 2010), index	110.0	12.34	0.421
Phonemic awareness; CTOPP, Ellison (Wagner et al. 1999), scaled score	11.83	2.5	−0.826
Orthographic ability; TOWRE, SWE (Torgesen et al. 1999), scaled score	109.97	10.56	−0.496
Decoding; TOWRE, PDE (Torgesen et al. 1999), scaled score	110.69	9.62	−0.849
Non-timed orthographic ability; WJ, Letter–Word (Woodcock and Johnson 1989), standard score	115.03	8.12	0.245
Reading comprehension; WJ (Woodcock and Johnson 1989), standard score	106.13	6.32	−0.499
Non-times phonological processing; WJ, Word Attack (Woodcock and Johnson 1989), standard score	109.52	8.84	0.530

TONI, Test of nonverbal intelligence (intact abilities >25); PPVT, Peabody picture vocabulary test (intact abilities >25); TOSREC, Test of silent reading efficiency and comprehension (average score 100 ± 15); CTOPP, Comprehensive test of phonological processing (average scores 10 ± 3); TOWRE, Test of word reading efficiency (average scores 100 ± 15); WJ, Woodcock Johnson (average scores 100 ± 15).

After scanning, study participants remembered an average of 63.27% (SD 17.28) of the words presented in the scanner task. The average d’ was 2.86 (SD 1.5), with a range of 0.9 to 5.4.

Correlation Between the Percentages of Words Remembered and Reading Measures

Comparisons between the percentages of words remembered (%hits) from the scanner task and the letter-word subtest (from the Woodcock Johnson battery) representing orthographical abilities revealed significant positive correlation between % words remembered and letter-word accuracy level (r = 0.451, P < 0.02), explaining roughly 20% of the variance. Individuals with a greater capacity for words remembered tended to have higher reading ability. Correlations with other reading measures were not significant.

The d’ accuracy measure was not significantly correlated with any of the reading measures.

Neuroimaging data: Robust activity resulted from word as well as checkerboard presentations, as expected (Fig. 1a), as well as for words alone (Fig. 1c). Differences in activation between words and checkerboards are shown in Figure 1b. Differences between remembered and forgotten words are shown in Figure 1d.

Figure 1.

Whole brain activation for (a) words and checkerboards (MET), (b) words vs checkerboards (Task × Time), (c) all words (Memory MET), and (d) remembered vs forgotten words (Memory × Time). All images have the same scale, showing voxels with z-scores above 2.5. Activations are displayed on an inflated brain using Workbench (Marcus et al. 2011).

Frames of data were removed from the GLM estimation if the frame-to-frame displacement was greater than 0.9 mm (Siegel et al. 2014). For the 27 participants in this analysis, an average of 9.7 frames were removed from the 263-frame run, with the worst participant result retaining 89% of the data.

Using our automated algorithm (https://readthedocs.org/projects/4dfp/), 43 regions that meet Monte Carlo multiple comparisons correction resulted from the MET image, 45 from Task × Time, 26 from Memory MET, and 10 from the Memory × Time image. Of these regions, 19 from the MET image had a significant Memory × Time effect, 17 from Task × Time, 4 from Memory MET, and all 10 from the Memory × Time image. After the removal of overlaps and regions with anomalous timecourses, 14 regions showing a difference between remembered and forgotten words remained, as shown in Figure 2 and listed in Table 2. Table 2 also shows which ANOVA image from which the regions were chosen.

Figure 2.

Regions chosen by automatic algorithm from the images in Figure 1 (a and c) and selected for showing a significant difference between remembered and forgotten words, as well as biological looking timecourses. Regions are colored by the effect shown in timecourses, and representative timecourses are shown for each effect. The top timecourses in each panel show activation for words in dark green and checkers in light green. The bottom timecourses show remembered words in dark blue and forgotten words in light blue. Effects shown are: gray—checker > word, remembered words positive, forgotten words flat; green – checker > word, remembered words positive, forgotten words negative; yellow—word = checker, remembered > forgotten; red - word > checker, remembered > forgotten. The box around the timecourse is the same color as the timecourse effect, and the arrows show which region the representative timecourse is from. Viewing conventions are the same as in Figure 1.

Table 2.

Statistics for timecourses in regions showing differences between processing of words subsequently remembered and those forgotten by typical child readers. All regions showed greater activation for remembered words than for forgotten words

Coordinates	Networks	Anatomy	Cubic mm	P-values
x,y,z source				Words and checkers	Words vs checkers	All words	R vs F	Effect
+37,−86,−04*	Visual	R inferior occipital gyrus	2160	7.49E-12	0.64	1.29E-07	2.04E-03	w = c
+28,−88,+20*	Visual	R middle occipital gyrus	1112	6.29E-07	4.24E-06	0.13	0.02	c > w r pos f flat
−20,−92,−04$	Visual	L inferior occipital gyrus	3080	6.83E-33	0.17	2.99E-29	1.34E-03	w = c
+24,−95,+02$	Visual	R middle occipital gyrus	2336	1.31E-23	0.47	2.14E-20	3.44E-05	w = c
−37,−81,−07$	Visual	L inferior occipital gyrus	2152	4.08E-17	0.01	2.77E-15	3.59E-03	w > c
+2,−83,+12*	Visual	R middle occipital gyrus	728	3.86E-07	3.75E-08	0.15	5.67E-04	c > w r pos f neg
+50,−11,+35*	Motor mouth	R precentral gyrus	1136	5.69E-07	1.12E-04	3.67E-03	6.94E-03	w > c
−28,−70,+23*	DA and Visual	L occipital gyrus	1512	2.09E-09	0.08	1.12E-03	0.03	w = c
+31,−73,+26*	DA and Visual	R occipital gyrus	544	2.78E-06	0.72	2.14E-03	0.02	w = c
−42,−04,+45*	DA, VA, FP, CO	L precentral gyrus	720	3.72E-06	0.73	3.81E-04	0.01	w = c
−54,−10,+37$	DA and Motor Mouth	L precentral gyrus	1904	1.39E-08	1.13E-05	2.46E-08	5.90E-03	w > c
−05,+07,+57$	CO, FP, VA	L middle frontal gyrus	1944	7.40E-10	0.03	5.50E-09	0.02	w > c
−37,+27,+03$	CO, VA	L inferior frontal gyrus	1280	2.01E-07	1.60E-04	1.69E-09	1.20E-03	w > c
+30,−69,−24*	Cerebellum	R cerebellum	976	1.56E-12	0.29	1.67E-04	3.30E-03	w = c

DA, dorsal attention; VA, ventral attention; CO, cingulo-opercular; FP, fronto-parietal; R, right; L, left; w, Words; c, Checkerboard; r, Remembered; f, Forgotten; pos, positive; neg, negative.

ANOVA image from which region was derived - *: Main Effect of Time – words vs checkers (all stimuli); $: Main Effect of Time – remembered vs forgotten (all words)

In addition to showing the locations of the 14 regions, Figure 2 also shows timecourses and location for the 4 patterns of activation observed. Words generated greater activation than checkerboards (red regions), mainly in the left hemisphere, while regions with equal activation for both stimuli (yellow) were more bilateral. Both regions showing greater activation for the checkerboard than the word presentations were in the right hemisphere. Regions showing a statistically significant difference in activation for remembered versus forgotten words were interpreted as those showing memory effects. All of these regions had greater activation for remembered words than those forgotten. In most regions, both remembered and forgotten words have a positive timecourse. However, in 2 regions in the right hemisphere, the timecourse for forgotten words is either flat (gray) or negative (green).

To determine which network the regions were in, we used previously published network boundaries (Power et al. 2011). Four of the 5 regions in the right hemisphere were in either Visual (3) or Mouth Motor (1) networks (Figure 3). The remaining right lateralized region crossed the border between Dorsal Attention and Visual networks (Figure 3), similar to the homotopic region in the left hemisphere. In the left hemisphere, by contrast, 2 of the 7 regions were in the Visual network and one was mainly in the Motor Mouth network. All other left hemisphere regions crossed the boundaries into 2 or more networks; Dorsal and Ventral Attention, Cingulo-opercular, Fronto-parietal, and Sensorimotor, as shown in Figures 3 and 4 and Table 2.

Figure 3.

Posterior regions along with network boundaries [following (Power et al. 2011)] along with posterior regions. Regions in the represented analysis include the Visual system (blue) and the Dorsal Attention network (green). Six of these regions are entirely within the Visual network, while 2 regions cross the Dorsal Attention to Visual network boundaries. Viewing conventions are the same as in Figure 1.

Figure 4.

Left hemisphere regions along with network boundaries (following (Power et al. 2011)] along with left hemisphere regions. Insets show details of 3 regions overlapping 2 or more networks. The top inset shows the boundaries both on top of and below the regions, since the regions cover much of the boundaries. The middle inset shows that the anterior insula/frontal operculum region crosses between the Ventral Attention and Cingular Opercular networks. The bottom inset shows the medial region is shared among 3 networks—Ventral Attention, Frontal Parietal, and Cingular Opercular. Viewing conventions are the same as in Figure 1.

During encoding, words subsequently remembered generated greater bilateral activation in visual processing regions (right cuneus, right middle occipital gyrus, left occipital gyrus) and left lateralized parts of the cognitive-control system (left medial frontal gyrus, precentral gyrus, inferior frontal gyrus and right cerebellum) (Table 2 and Fig. 2).

Discussion

The aim of the current study was to determine the neurobiological correlates of a subsequent word-memory task in children. Per our hypothesis and in line with previous studies in adults, we found higher activation of left visual-processing regions for words, both remembered vs forgotten. Interestingly, children, unlike adults in other studies, showed differential activation both in the left and right hemispheres in regions related to visual processing. Moreover, children also demonstrated higher activation of the cognitive-control systems in regions within the boundaries of networks related to alerting (ventral attention), orienting attention (dorsal attention), and executive control (fronto-parietal and cingulo-opercular) for remembered words compared to forgotten words. In support of theories related to the involvement of the cerebellum in reading and learning, we also found increased activation of the right cerebellum for words remembered than forgotten. Here, we discuss the role of the bilateral activation of the visual system (as opposed to left-lateralized engagement in adults) as well as cognitive-control systems including the cerebellum in reading and encoding words into memory in children.

Engagement of the Visual System in Encoding Words in Children: Left vs Right Hemispheres

Our study in children demonstrated an increased bilateral occipital activation for remembered words compared to forgotten words. This finding stands in contrast to previous results in studies of adults, which show that adults did not have activation in the right visual-processing regions during word reading retrieval (Wagner et al. 1998; Mei et al. 2010). However, a meta-analysis conducted in adults involving memory retrieval of linguistic stimuli (Kim 2011) suggested a role for the bilateral occipital regions in memory encoding for both pictures and words. Right hemisphere regions related to visual processing (right fusiform gyrus) were more active during pictorial stimulus presentation compared to verbal, presumably due to the need to process the picture in the right hemisphere. It is possible that children at the ages of 8–12 years who have mastered their reading, but are still constructing their mental lexicon, need to visualize both the word (in the left visual processing regions) and the corresponding “picture” of the item (in the right hemisphere) and therefore, both regions participate in this age group during word encoding. It would be interesting to examine whether younger children who have not yet mastered reading would have greater activation in the right hemisphere when requested to retrieve words from memory after reading them. Individuals with dyslexia may show a similar pattern of activation. For example, children whose reading is impaired (such as children with reading difficulties) are reported to have a reduced activation in the left fusiform gyrus and increased right fusiform gyrus during reading (Pugh et al. 2000). It may be that for children with reading difficulties, remembered words will cause more right lateralized occipital activation than in typical readers due to a heavier reliance on visualization of the corresponding picture of the word. A future study should look at this point in depth.

Involvement of Cognitive Control in Reading is Associated with Remembered Words in Children

The results of the current study echo previous studies of subsequent memory in adults, which demonstrate a greater activation for remembered words than forgotten words in the left frontal regions; see (Buckner et al. 1999) and (Wagner et al. 1998). In line with that, Kim also showed greater frontal activation for verbal information remembered than forgotten and suggested that both the left lateralized regions found in the meta-analysis (left inferior frontal cortex and bilateral fusiform cortex) are canonical content-processing regions, critical for encoding verbal information (Kim 2011). Buckner and colleagues suggested that the left dorsal frontal regions participate in the process of remembering words (Buckner et al. 1999). This review proposed that only when frontal regions are involved in a cognitive process would the temporal lobe be capable of encoding the information it processed.

It is obvious that during the reading process, the involvement of the visual system is important for encoding the words being read. An examination of functional connectivity patterns between visual and cognitive-control systems would elucidate the relationship between these 2 activations. In line with our hypothesis, the regions more active for remembered words in the current subsequent memory task fall within the boundaries of several key networks described in Petersen and Posner’s model for cognitive control (Petersen and Posner 2012). As suggested in the model, when individuals were asked to attend to words and to remember them for a test after the scan, their alertness to the external stimuli rises (i.e., ventral attention network) and they orient their attention to the words (i.e., dorsal attention), at least for those that were remembered. Also, remembered words causing greater activation within networks related to internal executive-control processes such as monitoring and information processing occurred (i.e., cingulo-opercular and fronto-parietal networks); see (Petersen and Posner 2012). This greater activation for remembered words is in line with the increased P300 and P600 components found in adults (Paller et al. 1987). An interesting next step would be to conduct a task-based functional connectivity analysis using these networks for words remembered vs forgotten. Such an analysis has the potential to demonstrate the changes in functional connectivity between networks related to alerting and orienting attention, as well as executive-control networks, with time.

The cerebellum has been well described in imaging-based studies of language processing and reading (Vlachos et al. 2007; Shaywitz and Shaywitz 2008; Berl et al. 2010; Price 2012; Hutton et al. 2017). Associations between cerebellar microstructure and reading-component skills in school-age children have also been described (Travis et al. 2015). It has been proposed that the cerebellum applies a “cerebellar transform” (Guell et al. 2015) to higher-order cognitive processes, enhancing skill refinement and learning (Ito 2008; Buckner 2013). Cerebellar activation seems to increase with greater cognitive loading, such as new or challenging words (Marien et al. 2014), particularly involving higher working memory demands (Kuper et al. 2016). Interestingly enough, the role of the cerebellum in memory formation was previously suggested when adults were presented with pictures and had to recognize which picture they saw while being scanned (Weis et al. 2004). Correctly recognized pictures were positively associated with increased cerebellar activation. In the context of the current study results, these findings seem reasonable given evidence for a cerebellar role in visual attention and working memory (Brissenden et al. 2016), which are core executive skills supporting reading development.

One additional intriguing finding of our study was Mouth Motor network activation during silent word reading. We suggest that even the silent reading of words may involve mouthing of words as a strategy in children, which may also assist in consolidation of words in memory. Additional studies of children in different stages of reading acquisition and with different reading abilities (i.e., children with dyslexia vs typical readers) should support the relationship between the ability to remember words and reading ability.

Study Limitations

The current study has its limitations: First, since there are no behavioral measures for passively viewing words, it is possible that words that were not remembered simply were not attended to. Second, in order to better understand the relationships among visual, dorsal and ventral attention, frontal parietal and cingulo-opercular systems, a functional connectivity analysis should be performed. Unfortunately, we did not collect enough resting-state data to look at individual effects, so we cannot determine whether our regions are at the junctions of different functional networks because of the involvement of multiple networks (or the regions connecting them) or primarily due to individual variation in the locations of these networks. Third, our behavioral test for which words a participant remembered were binary and did not measure the degree of certainty a participant felt in the memory. Another intriguing question is whether the relationship of semantics to an increased subsequent memory effect previously reported (Paller et al. 1987) exists in our study. Despite the high-frequency words used in the current study, our participants were children and may have not known the meaning to some of the presented words. Since we did not ask the children for the meaning of the words they forgot, we cannot confirm this hypothesis. Future studies should look at this point, as well as examine this question in children with lower reading ability, such as in children with dyslexia, who may not recognize some of the words (i.e., do not have a semantic representation for them).

Conclusions

The current study is the first in children to distinguish between the involvement of the visual and cognitive-control systems during reading remembered vs forgotten words. As was previously found in adults for verbalizable stimuli such as pictures or words, bilateral activation in visual-processing regions and left lateralized frontal regions was observed. The addition of the cerebellum in this study, which was not observed in previous studies in adults, may reflect this task’s additional challenge for children in the early phases of mastering reading. In light of the accumulated data demonstrating an alteration in both the visual (i.e., left fusiform gyrus) and cognitive-control systems (i.e., anterior cingulate cortex and dorsolateral prefrontal cortex) in children with reading difficulties (Horowitz-Kraus and Holland 2015; Horowitz-Kraus, DiFrancesco et al. 2015; Horowitz-Kraus, Toro-Serey et al. 2015), we postulate that some of the challenges this population has in retrieving words from memory are related to their inability to activate the above-mentioned regions appropriately during reading. A future study examining the involvement of visual and cognitive-control systems during reading in children with dyslexia, and also while retrieving written words from memory, is critical to verify this point.

Highlights

• Greater activation in left frontal and bilateral occipital regions for word reading vs checkerboard viewing in children

• Greater bilateral activation in visual processing regions for words remembered vs forgotten in children

• Location of regions with greater activity for words subsequently remembered on borders between functional networks

• Greater left lateralized activation in cognitive-control regions for words remembered vs forgotten in children

• Involvement of the cerebellum in word reading

Funding

This work was supported by the National Institutes of Health (R01 HD086011) to T.H-K.