Abstract
Human neuroimaging studies of episodic memory retrieval routinely observe the engagement of specific cortical regions beyond the medial temporal lobe. Of these, medial parietal cortex (MPC) is of particular interest given its distinct functional characteristics during different retrieval tasks. Specifically, while recognition and autobiographical recall tasks are both used to probe episodic retrieval, these paradigms consistently drive distinct spatial patterns of response within MPC. However, other studies have emphasized alternate MPC functional dissociations in terms of brain network connectivity profiles or stimulus category selectivity. As the unique contributions of MPC to episodic memory remain unclear, adjudicating between these different accounts can provide better consensus regarding MPC function. Therefore, we used a precision-neuroimaging dataset (7T functional magnetic resonance imaging) to examine how MPC regions are differentially engaged during recognition memory and how these task-related dissociations may also reflect distinct connectivity and stimulus category functional profiles. We observed interleaved, though spatially distinct, subregions of MPC where responses were sensitive to either recognition decisions or the semantic representation of stimuli. In addition, this dissociation was further accentuated by functional subregions displaying distinct profiles of connectivity with the hippocampus during task and rest. Finally, we show that recent observations of dissociable person and place selectivity within the MPC reflect category-specific responses from within identified semantic regions that are sensitive to mnemonic demands. Together, by examining precision functional mapping within individuals, these data suggest that previously distinct observations of functional dissociation within MPC conform to a common principle of organization throughout hippocampal–neocortical memory systems.
Significance Statement
The medial parietal cortex (MPC) is ubiquitously implicated in studies of human episodic memory retrieval. Interestingly, distinct subregions of MPC are engaged depending on the type of retrieval task being performed. Other types of functional dissociation have also been reported for MPC, focusing on brain network organization or stimulus selectivity. To reconcile these different functional views, we leveraged a precision-neuroimaging dataset where participants performed thousands of recognition memory trials. We observed distinct MPC subregions supporting either recognition decisions or their semantic content, which recapitulated distinct patterns of hippocampal connectivity and visual category selectivity within MPC. These findings promote a common principle of functional organization that is shared among brain regions supporting episodic memory and advances our understanding of MPC.
Introduction
The medial parietal cortex (MPC; Fig. 1a) is routinely observed in neuroimaging studies of episodic memory retrieval (Wagner et al., 2005; Vincent et al., 2006; Buckner et al., 2008; Daselaar et al., 2009; Kim and Cabeza, 2009; Huijbers et al., 2011), displaying larger magnitude responses with increasing memory strength (Bird et al., 2015; Ritchey et al., 2015) and strong anatomical connectivity with medial temporal lobe structures (Vogt and Pandya, 1987; Parvizi et al., 2006). However, despite the ubiquitous observation of MPC activity during episodic retrieval, little is known about this region's unique contributions to mnemonic processing and therefore its role in cognition more broadly (Foster et al., 2023). Traditionally, investigators have used tasks requiring the recall of autobiographical events, or more commonly, laboratory-based tasks of studied item recognition, to experimentally probe episodic memory (Cabeza and St Jacques, 2007; Rhodes et al., 2019). Strikingly, it has long been observed that despite both kinds of task being viewed as assays for episodic retrieval, the MPC displays distinct spatial patterns of response to each of these tasks (McDermott et al., 2009; H-Y. Chen et al., 2017; Fig. 1b). Gaining a better understanding of this marked functional dissociation is important not only for characterizing the role of the MPC in mnemonic processes but also for better understanding the mechanisms that support distinct cognitive stages and demands during recall and recognition.
MPC anatomy, NSD experimental paradigm, and analysis approach. a, MPC reflects the medial aspect of the parietal lobe, also termed posteromedial cortex (PMC). Demarcation of the MPC (green) is shown on a standard (left) and inflated (right) cortical surface. b, MPC contains clear cytoarchitectural subdivisions, comprising the PrC, PCC and RSC (left). Schematic (right) depicts common dissociable spatial patterns observed within the MPC by fMRI studies of recognition (red) and autobiographical (blue) tasks of episodic retrieval. c, During the NSD stimulus-recognition task, subjects were shown one image at a time (3 s) and asked to indicate whether the image was “old” or “new”, followed by a brief period where the fixation dot remained on the screen (1 s). d, The NSD includes data from eight subjects as they performed the stimulus-recognition task over the course of multiple sessions. Our analyses focus on a subset of those sessions (Sessions 1–12), which included memory decisions spanning up to 100 d since the time of encoding (9,000 trials). e, Stimuli are from the COCO image dataset, which includes detailed image content segmentation and associated text annotations. These text annotations were used to perform semantic similarity analysis across stimuli (see Materials and Methods). f, Data analyses focused on identifying MPC responses sensitive to the mnemonic status (recognition regions) and semantic content (semantic regions) of presented stimuli. Given the focus on single-subject-level analysis, we performed an odd/even split of experimental scanning blocks to cross-validate all core findings.
A growing literature suggests that the observed dissociation of MPC responses during different types of episodic memory tasks may be better understood by considering the distinct types of mnemonic demands they require. For example, retrieval and representation of memory content predominates during autobiographical recollection, while control processes necessary for memory-based decisions are essential for recognition memory (Leech et al., 2011; Bastin et al., 2019; Aponik-Gremillion et al., 2022). Consistent with this view are observations that item-recognition responses in the MPC are sensitive to both the control demands of tasks and mnemonic demands (Rosen et al., 2016, 2018; Gilmore et al., 2019a; Isenburg et al., 2023). Furthermore, regions implicated in recognition overlap with the cognitive control (Vincent et al., 2008; Yeo et al., 2011; Willbrand et al., 2022) or parietal memory (Gilmore et al., 2015) networks. Conversely, MPC regions linked to autobiographical retrieval primarily overlap with the default mode network (DMN; Spreng and Grady, 2010; Andrews-Hanna et al., 2014). This fractionation is consistent with the heterogenous anatomy of the MPC (Bzdok et al., 2015; Ritchey and Cooper, 2020; Dadario and Sughrue, 2023; Rolls et al., 2023), which comprises multiple subregions, specifically the precuneus (PrC), posterior cingulate cortex (PCC), and retrosplenial cortex (RSC; Foster et al., 2023; Fig. 1b). However, the considerable individual variation in functional and anatomical organization found in associative cortical regions such as the MPC (Margulies et al., 2009; Vogt and Palomero-Gallagher, 2012) makes standard neuroimaging methods that rely on spatial group averaging suboptimal for the detailed characterization of this region's functional neuroanatomy (Braga et al., 2019).
Recent developments in precision-neuroimaging methods, where a small number of subjects are densely sampled through repeated scanning, not only improve the specificity of functional mapping within individuals but also provide increased statistical power for conducting previously unfeasible analyses (Laumann et al., 2015; Gordon et al., 2017b, 2023; Gratton and Braga, 2021). To this end, we analyzed an openly available functional magnetic resonance imaging (fMRI) dataset (Allen et al., 2022) collected at high resolution using an ultrahigh-field scanner (7T) while eight subjects performed thousands of stimulus-recognition decisions about semantically rich naturalistic images over the course of many experimental sessions. This massive dataset afforded the statistical power and anatomic precision for evaluating MPC responses during item-recognition task performance that either reflected involvement in memory-based decision-making or were more indicative of mnemonic content representation. We observed interleaved, though distinct, functional subregions of the MPC where responses were sensitive to either recognition decisions or representation of the semantic content of stimuli, but rarely both. This striking functional dissociation was recapitulated in functional connectivity differences with the hippocampus and patterns of categorically selective responses to person or place stimuli during recognition task performance. These results help to better account for how distinct patterns of response may occur as a result of task demands during episodic retrieval and may reflect a common principle of organization throughout hippocampal–neocortical memory systems.
Materials and Methods
The data analyzed here were collected and initially preprocessed as part of the openly available Natural Scenes Dataset (NSD; Allen et al., 2022). The NSD is a large-scale fMRI dataset collected using ultrahigh-field fMRI (7 T) while subjects completed hundreds of stimulus-recognition decisions on semantically rich naturalistic images (from the COCO dataset; Lin et al., 2015). The NSD includes preprocessed, high-resolution functional images (beta-maps with upsampled 1 mm isotropic voxels and robust HRF fitting), but also fMRI data from rest and visual localizer tasks. For a full description of the dataset, as well as specific methods involved in collecting and preprocessing the original data, please see the original manuscript (Allen et al., 2022). Below, we outline the analysis steps relevant to the current manuscript.
Subjects
Eight subjects from the University of Minnesota community were included in the experiment (two male and six female, ages 19–32). All subjects had normal or corrected to normal vision, and no known cognitive deficits. Informed written consent was collected for all subjects, and the protocol was approved by the University of Minnesota institutional review board.
MRI acquisition and preprocessing
Anatomical data was collected using a 3T Siemens Prisma scanner and 32-channel head coil. For the analyses included in this manuscript, we used the preprocessed, 1 mm preparations of the volumetric anatomical images and the corresponding FreeSurfer output for surface spaces. Additionally, we used the provided transformation matrices and code to map between individual volumetric and surface spaces, as well as to normalize functional outputs from individual surface space to the fsaverage surface. Task and resting-state functional scans were collected using a 7T Siemens Magnetom passively shielded scanner using a 32-channel head coil. For functional analyses, we used the 1.0 mm preparation of the functional timeseries data (described as beta-2 in the original manuscript). Briefly, this preparation is characterized by treating each trial as an independent “condition” and fitting each voxel with a best-fitting hemodynamic response function (HRF) determined for each voxel and session. This is done using a cross-validated approach for fitting a library of potential HRFs to each voxel, choosing the function with the best fit, and using that function to evaluate the entire session for that voxel. Next, these preprocessed HRF-fit volumetric timeseries were resampled by cubic interpolation onto the subject-native surface at three different depths (layers 1–3) and then averaged across all depths to create a single surface timeseries matrix (vertices × trials). These matrices were used for all subsequent functional analyses except for hippocampal timeseries data, which was taken from the corresponding volumetric timeseries (presurface interpolation voxel × trial matrices). No smoothing was applied to these timeseries. To return data from int16 minimal-storage format to percent signal change for analyses, all vertex values were divided by 300. Finally, values were normalized by z-scoring within each vertex and session.
Task details
Subjects performed a continuous image recognition task, where on each trial, they were shown an individual image and asked to respond as to whether the image was identical to any image that had been presented previously during an experimental session (“old image”) or had not been presented during any experimental session (“new image”). The experiment was organized such that each scanning session contained 12 runs, with 75 trials presented each run. The first three and last four trials of each run were blank, with five additional blank trials being sporadically presented within each run (time between blank trials 9–14 trials). Additionally, the 63rd trial of even runs was blank, leading to a total of 750 task trials per session [(63 trials * 6 odd runs) + (62 trials * 6 even runs) = 750]. Subjects completed between 30 and 40 scanning sessions each, taking part in experimental sessions approximately once a week over the course of a year. Depending on the number of sessions completed over the course of the entire experiment, subjects performed between 22,500 and 30,000 trials. Each subject was shown up to 10,000 unique images, and each image was repeated up to three times. Across all subjects, 73,000 unique stimuli were employed, with 1,000 stimuli chosen to overlap between all eight subjects.
The NSD stimulus-recognition task was designed so that individuals could continuously make memory decisions on hundreds of visual stimuli across a wide range of inter-repetition gaps. Each trial lasted for 4 s, wherein a single stimulus and a small, semitransparent fixation dot were presented for 3 s, followed by a 1 s interstimulus interval during which only the fixation dot remained on the screen. Subjects were instructed to maintain central fixation and to respond by pressing 1 with their index finger for “new” images or pressing 2 with their middle finger for “old” images. Subjects were allowed to change their response as many times as they wanted during the trial window, and accuracy was defined by their last button press. “Old” trials could be those where stimuli had been shown either previously during the current session (“easy-old” trials) or during a previous session (“hard-old” trials). The task was designed so that as subjects performed the experimental sessions over time, the number of old items increased, the ratio of easy to hard trials decreased, and the ratio of new to old trials decreased. As such, a cutoff of trials to be included in subsequent functional analyses was chosen such that the ratio of old to new trials, as well as the ratio of hard to easy trials, remained below 1.5:1. Importantly, the shift in ratio of trial types corresponds to a similar drop in recognition performance. We evaluated adjusted hit rates for easy- and hard-old trials (the rate of correct recognition decisions for easy or hard trials divided by the overall false alarm rate; Allen et al., 2022) for sessions which all subjects performed and were publicly available (Sessions 1–26). Performance was significantly worse for both easy- and hard-old trials after the 12th experimental session (easy: t(199) = −10.71, b = −0.140, 95% CI = [−0.165, −0.113], p < 0.001; hard: t(191) = −10.995, b = −0.109, 95% CI = [−0.129, −0.090], p < 0.001). This cutoff, set at including Sessions 1–12, considered both the trial-type ratio and subject performance and was chosen in order to maximize generalizability of the results from the current analyses while still taking advantage of the statistical power to evaluate memory decisions afforded by the NSD task design.
Analysis details
Recognition decision and semantic content analysis overview
Our first set of analyses were aimed at identifying regions of the MPC where activity was associated with recognition memory decisions and/or semantic content. To ensure robust results, a cross-validation sampling method was used. The z-scored cortical surface data was first divided into even and odd runs from the first 12 fMRI sessions. Then recognition decision and semantic content analyses were performed to separately create clusters on one set of the data (e.g., even trials) and to test cluster analysis scores on the held-out set (e.g., odd trials). This process was repeated once, so that each partition of data served once as the cluster-generating dataset and once as the score-generating dataset. All analyses were performed in Python 3.7 using custom code from NiLearn, scikit-learn (Pedregosa et al., 2011), NiBabel (Brett et al., 2022), and SciPy (Virtanen et al., 2020) packages, as well as code provided by the NSD research team (https://github.com/cvnlab/nsdcode).
Recognition decision analysis
In order to identify regions of the MPC where activity was related to recognition decisions, we first performed a univariate analysis of blood oxygenation level-dependent (BOLD) activity during hits versus correct rejections. At the first level of this analysis, the normalized beta values within each session for each subject were compared between hit recognition trials (correct “old” responses) and correct rejection trials (correct “new” responses), generating a by-session t-score map. Next, we tested whether differences in activity between hits and correct rejections were consistently greater than would be expected by chance by running a one-sample t test comparing the by-session t-scores with an empirical mean of 0 for each vertex. The corresponding t-scores from this second-level analysis were normalized (to z-scores) for thresholding and comparison across subjects. In order to control for multiple comparisons without drastically decreasing analysis sensitivity, cluster thresholding was performed by including only clusters where at least 20 contiguous vertices were above a z > 3.3 (p < 0.0005) threshold. For one subject (S8), a threshold of z > 3.3 led to no surviving clusters. For this subject, the threshold was lowered to z > 3.1 (p < 0.001) so that we could run a more complete statistical analysis of reliability across individuals. These thresholded maps were used for the visualization and characterization of recognition responsive regions within the MPC.
Semantic content analysis
We also sought to identify regions of the MPC where BOLD activity was related to the semantic content of recognition stimuli. In summary, semantic content regions were identified as those where activity patterns during recognition were correlated to a pretrained semantic model applied to each image (Google USE_v5; Cer et al., 2018). This analysis involved a two-component surface searchlight approach. First, for each stimulus, a semantic feature vector was generated from the annotation captions included for each image in the COCO dataset (more details on the caption collection and validation can be found in the original manuscript; X. Chen et al., 2015). The annotations were five unique human-generated sentences describing each stimulus (e.g., “a boy in a gray sweater holding a wooden baseball bat”). These sentences were converted to quantitative semantic feature vectors for each trial, by concatenating all five sentences and feeding them into Google's Universal Sentence Encoder (USE). The Google USE_v5 encoding model was selected as it is optimized for multiword semantic evaluation. The result of feeding annotations through the semantic model was a 512-semantic feature vector for each stimulus. The dissimilarity between each stimulus's semantic vectors was measured using cosine distance. This generated a stimulus × stimulus-sized matrix of semantic dissimilarity for each stimulus seen during the included trials (e.g., even/odd runs) of each session.
The second step of the semantic content analysis involved generating a corresponding dissimilarity matrix for stimuli based on BOLD patterns of activity. A surface searchlight approach was used to generate these neural dissimilarity matrices. For each vertex on the surface, a 3 mm radius searchlight sphere was generated (mean vertices = 90) and converted into a vector. The cosine distance of BOLD activity in response to each stimulus was calculated to generate a vertex-centered dissimilarity matrix for each searchlight sphere across all included trials. Lastly, the relationship between the neural dissimilarity matrix and the semantic model's dissimilarity matrix was quantified by taking Pearson’s correlation between the two matrices. The output correlation value was assigned to the sphere center as the semantic similarity score for that vertex. This process was repeated for all vertices, across all sessions, and within each subject in the same cross-validated manner as above. Having computed the by-vertex, within-session correlations between BOLD activity and the semantic model, we next sought to determine how reliable this relationship was across sessions. The quantification of semantic content reliability was done by normalizing correlation values (Fisher transforming) and then computing a single sample t-score to evaluate statistical difference from zero for correlations across sessions. Next, the resulting semantic reliability maps (t-scores) were normalized (to z-scores), and a cluster thresholding of z > 3.3 (equivalent to p < 0.0005) and a cluster size minimum of 20 contiguous vertices were applied. In order to mirror the memory-decision analysis, this cluster threshold value was lowered to z > 3.1 (p < 0.001) for S8 (the same subject as above).
Next, we investigated group-level results for both the recognition and semantic content analyses so that we could compare findings to previous research. Here, prethresholded second-level z-score maps from each analysis for each subject were normalized to the fsaverage surface using the “NSD_mapdata” function and publicly available NSD transformation matrices. Group-map t-scores were computed within each vertex and then converted to normalized z-scores for thresholding and visualization.
To compare correlations between the Google USE output and responses across behavioral trial types, we separately performed the cross-matrix correlations for only hit or correct-rejection trials within each session, via the same semantic content reliability analysis as noted above. Lastly, voxels were deemed to overlap between analyses if they appeared in clusters in both the hit and correct-rejection analyses. To evaluate the strength of correlation responses, we submitted the average z-scores found within hit and correct-rejection semantic clusters for each subject to a paired t test.
A cluster-based analysis was next carried out to quantify how functional responses differed between MPC clusters implicated in the recognition decision and semantic content analyses within individuals. Using the cross-validation method outlined above, clusters were generated using either even or odd runs across the first 12 sessions. Then the held-out set of runs was used to generate score maps for each subject (e.g., strength of recognition or semantic content responses). This scheme was repeated once so that each set of data was used once to generate clusters and once to generate z-score maps. To check whether analysis scores were consistent across the two cross-validated data-folds, we used mixed effects linear regression with the difference in cluster scores between even and odd runs as the fixed effect and a random effect of subject intercept. Having established scores did not significantly differ across folds; we averaged results across folds and then quantified the degree of overlap between regions implicated in the recognition versus semantic content analyses. To do so, we took the four output values for each subject, representing a factorial combination of cluster type (recognition or semantic content) by analysis map (recognition or semantic content), and input them into a mixed effects linear regression model that also included a random subject intercept term and fixed effects of cluster type and analysis type. Results from this analysis are visualized in Figure 2c.
Stimulus-recognition and semantic-based responses within the MPC. a, Regions associated with recognition decisions (red), image semantic content (blue), or both (green) visualized on the inflated medial surface anatomy for each individual. Depicted clusters were thresholded at z > 3.3 (p < 0.0005) for all subjects except for S8 (z > 3.1). b, Group map for the same contrasts and threshold as (a). c, Average recognition (left) and semantic (right) responses measured from recognition clusters (red) and semantic clusters (blue) from each subject. Scores were consistent over cross-validation data-folds (even/odd). Left: Normalized BOLD responses to recognition decisions were significantly greater in recognition clusters than in semantic clusters. Right: Semantic responses were significantly greater in semantic clusters than in recognition clusters. Colors and dashed lines represent individual subjects. d, Mean percentage of cluster type (recognition, semantic, or both) assigned to MPC surface vertices (both hemispheres) at the group level and for each individual.
To better characterize the relationship between recognition and semantic responses in MPC, we averaged the z-score output from the across-session analyses for each vertex across data-folds. This results in one recognition and one semantic value for each vertex, which we then submitted to a correlation for all MPC vertices. To test for a consistent correlation within regions of interest (ROIs) across subjects, correlation values were first Fisher-z transformed and then submitted to an across-subject t test against a null of zero.
We also evaluated the degree of overlap between recognition and semantic content regions by quantifying the percentage of MPC vertices that we observed to be above threshold for the recognition, semantic content, or both analyses. This value was determined by taking the average number of vertices that were above threshold and then dividing by the total number of MPC vertices (visualized in Fig. 2d). To ensure that results were robust to cluster parameter selection, we performed cluster identification by threshold free cluster enhancement (TFCE) and repeated the overlap analyses reported above. Results from the TFCE (FWE corrected, p < 0.05) and traditional cluster thresholding were qualitatively similar, and as such only the original, primary analysis is reported in detail here. These analyses allowed for us to gain insight into the degree and arrangement of processing overlap for recognition and semantic content within MPC subregions.
To evaluate the potential relationships between behavior and BOLD responses, we performed two further analyses. First, we correlated responses (% signal change) on each trial to RT within each MPC vertex with each individual session. Next, we Fisher-z transformed these within-session correlations and compared the distribution of correlations with a null correlation score of zero. This generated a t-score indicative of the reliability of correlations across sessions within each vertex. Finally, we averaged these t-scores across recognition versus semantic clusters for paired t test comparison. We next performed an analysis to determine if responses in recognition versus semantic clusters differed across trial types (hit–easy-old, hit–hard-old, correct-rejection–novel). To do so, we first computed the within-session t-score for response strength to each trial type across all vertices. Next, we performed an across-session t test against zero to evaluate the strength of responses for each trial type. We then normalized the output of these across-session t tests (z-scores) and took the average from within recognition and semantic clusters for each trial type. Lastly, we fit both linear and quadratic models to the across-subject data in order to evaluate the relationship of responses across trial types.
Category functional localizer (fLoc) experimental analysis
The fLoc (developed by Stigliani et al. (2015); http://vpnl.stanford.edu/fLoc/) was used to examine responses to isolated visual object categories. This fLoc has been previously used to identify regions involved in face or place processing (Stigliani et al., 2015) and involved displaying curated images from 10 categories to subjects. The stimulus categories were either characters (words or numbers), bodies (bodies or limbs), faces (adult or child), places (houses or corridors), or objects (cars or instruments). The images used in these stimuli have been carefully selected and cropped to display only features in the category of interest. Stimuli were presented in grayscale on scrambled backgrounds and adapted to fill an 8.4° × 8.4° square and with a semitransparent fixation dot centrally located, mirroring the NSD stimulus presentations. Stimuli were displayed in eight trial miniblocks with each stimulus being presented for 0.5 s each, totaling 4 s per miniblock. Each run contained six presentations of each of the 10 categories, with each run lasting 300 s total. Six runs were collected, totaling 36 miniblocks and 288 image presentations per category.
The NSD includes t-score contrasts from the category localizer task, reflecting categorical selectivity across individual cortical surfaces. These selectivity maps were generated using GLMdenoise and a “condition-split” strategy (for more details, see the original manuscript and supplementary files). For person selectivity, t-scores were generated by comparing betas for face stimuli (adult and child) to all other categories. Similarly, place selectivity was computed by comparing house and corridor responses with all other categories. We then generated selective clusters by taking these precomputed t-score maps, thresholded at t > 3.3, and applied a cluster threshold of necessitating at least 20 contiguous vertices. These clusters were then applied to the NSD-based categorical selectivity analysis described below. Additionally, from each of the six localizer runs, a series of beta values for each vertex were made available as part of the NSD. We evaluated these beta values to determine the univariate direction of person and place responses, in order to compare the current findings with previous work (Silson et al., 2019). It is important to note that while the fLoc task had a similar number of trials compared with previous studies that have included perceptual localizers, it had far fewer trials than the NSD stimulus-recognition task, making the statistical power different between the fLoc and recognition task analyses.
Categorical representation analysis
While much of previous research examining categorical selectivity has used experiment-controlled stimuli, such as the fLoc above, the NSD's use of the COCO stimuli allows for examination of person- and place-selective processing in the MPC across a wide range of naturalistic images. Our first step in performing a categorical selectivity analysis was to sort NSD stimuli into those that primarily featured either people or places. We used pregenerated bounding boxes (https://github.com/cocodataset/cocoapi and https://github.com/ACI-Institute/faces4coco) to establish the percentage of each image that was occupied by a person's body or face, or by any objects or animals within the stimulus. If at least 10% of the image was a face or 15% of the image was a person's body (face, body, or limbs), we considered that image to be highlighting an individual and to be a “person” image. Conversely, to select “place” images, we performed the opposite inclusion steps. If the stimulus included at least 15% of any given object (animal, food, etc.) or any amount of a person, we determined that the stimulus featured an object, person, or animal and as such removed that stimulus from the “place” image set. All other stimuli were included. This stimulus selection process allowed for a robust examination of categorical selectivity within the MPC by including complex naturalistic stimuli that featured either people or places, but also contained other objects or actions. Examples of included and excluded stimuli can be seen in Figure 4a.
Next, we quantified categorical selectivity within the MPC by evaluating the same preprocessed data as used in the previous recognition and semantic content analyses (the beta version-2 timeseries surface data from Sessions 1–12 of the NSD recognition task). In a similar fashion to the previous recognition and semantic analyses, we conducted this analysis in a cross-validated manner, whereby one half of the data (e.g., even runs) was used for analyses and then the process repeated with the held-out set of data. By-session bias scores for either person or place stimuli (t-scores) were generated by comparing betas from our sorted person with place stimuli for correct trials only (hits and CRs). Next, the by-vertex statistics of categorical bias were quantified by conducting a single-sample t test to compare the bias score across sessions with the empirical chance of zero. Across session categorical selectivity bias maps (t-scores) were normalized to z-scores for thresholding and visualization. For cluster generation, cluster thresholding was performed to the second-level categorical selectivity maps from the NSD task (z-scores) analysis to only include vertices where z > 3.3 (p < 0.0005) and which were part of a group of at least 20 contiguous vertices. For visualizing patterns of categorical selectivity in the MPC within individuals (Fig. 4b), selectivity maps from each data-fold were averaged together, and thresholding was applied to those averaged maps.
To confirm that this analysis was capturing similar aspects of stimulus processing as the well-controlled category localizer, we compared results from the localizer and NSD-task categorical analyses. This involved using the person- and place-selective clusters generated from the category localizer task as masks to evaluate the person/place selectivity from the NSD-task selectivity analysis. To ensure the robustness of results, this process was repeated across each data-fold of the NSD recognition data, such that z-score categorical selectivity maps were generated once from even and once from odd runs. Categorical selectivity scores within category localizer-generated clusters did not differ across data-folds (t(22) = 0.049; b = 0.012; 95% CI = [−0.504, 0.528]; p = 0.961), confirming the consistency of results and allowing for scores to be averaged across folds for further analysis. Significant selectivity in the same direction as the category localizer clusters was next evaluated as evidence that the NSD-task selectivity analysis was capturing similar elements of categorically biased responses during recognition memory performance.
Next, we sought to compare how regions demonstrating categorical selectivity related to networks of regions implicated in either recognition or semantic content. To do so, the person and place clusters from the NSD task were used as masks for generating mean responses from the recognition and semantic content analyses. We used cross-validation to generate clusters from one half of the data (e.g., even runs) and generated z-scores from the held-out set (e.g., odd runs). Responses were averaged across data-folds and then were compared using mixed effects linear regression, including random subject intercepts and fixed effects of contrast type (recognition/semantic) and cluster type (person/place). As an additional means of measuring overlap between functional regions of the MPC, we quantified the percentage of categorically selective MPC vertices that were also identified as belonging to recognition or semantic clusters. For one subject (S6), no person clusters survived thresholding, so they were excluded from comparisons between person and place clusters. In order to evaluate the robustness of results to clustering threshold parameters, we also applied TFCE to the person/place contrast. Results from the TFCE (FWE corrected, p < 0.05) and traditional cluster thresholding were qualitatively similar, and as such only the original, primary analysis is reported in detail here.
The NSD task contains greater than an order of magnitude more person and place trials than the fLoc task. Therefore, to better control for statistical power between tasks for direct comparison, we also subsampled the NSD task and analyzed the strength of categorically selective responses during each individual session independently. To generate ROIs for this comparison, we used the across-session person and place clusters that had previously been identified and used those as masks for averaging categorically selective responses from each NSD session and for the fLoc task. We performed an across-subject paired t test to compare whether the average categorical responses from the NSD task were significantly different from the fLoc task. Lastly, to evaluate how responses in the MPC compared with a nearby region regularly implicated in perceptual place selectivity, we performed the same analysis within the ventral parieto-occipital sulcus.
We also performed additional analyses aimed at further characterizing person- and place-selective MPC regions. First, connectivity with the ventral temporal cortex was measured by isolating the parahippocampal place area (PPA) and fusiform face area (FFA) for each participant as identified from the category localizer (fLoc task) with liberally thresholded masks (t > 1) included in the NSD (see the original manuscript for more details). We then correlated the beta timeseries from each session between person and place MPC clusters with the ventral temporal PPA and FFA regions. These correlations were normalized (using the Fisher-z transform), and then the difference in correlation between the PPA and FFA was calculated for each MPC vertex, giving a by-session connectivity bias measure. This difference measure was calculated for each session, and then the session scores were submitted to a one-sample t test and normalized to give across-session connectivity bias scores (z-score). This output was collapsed across hemispheres and submitted to an across-subject mixed effects linear regression. Additionally, we tested for differential activity during recognition decisions. To do so, we took the average difference between responses to hits versus correct rejections (z-scores) from within person- and place-selective MPC regions for each subject and then submitted those values to a paired t test.
Functional beta-series correlation analysis
We also sought to test whether MPC regions involved in recognition decisions or semantic content may be differentially linked to anterior versus posterior hippocampal subregions. We used functional connectivity between hippocampal and MPC subregions to evaluate this relationship. Individual-specific anterior and posterior hippocampal ROIs were designated by manually dividing hippocampi at the uncal apex (Poppenk et al., 2013). Task-based MPC to hippocampal subregion connectivity bias was quantified by looking at differences in beta-timeseries correlations between the MPC and anterior versus posterior hippocampal voxels during NSD recognition task performance (Rissman et al., 2004). The normalized volumetric data (z-scored 1.0 mm, beta version-2) was first correlated between all MPC and hippocampal voxels separately within each subject and session. For each session, these MPC–hippocampus correlation maps were Fisher transformed, and then the average MPC correlation with the anterior hippocampus and with the posterior hippocampus was calculated. Lastly, MPC–hippocampus bias maps for each session were created by subtracting the posterior average from the anterior average for each MPC voxel. This MPC–hippocampal bias score was computed separately for voxels residing in each hemisphere (e.g., left MPC to left hippocampus, left MPC to right hippocampus, etc.). Hippocampal bias scores were input into a second-level across-session t test, which determined whether MPC voxels demonstrated a systematic anterior or posterior connectivity bias during task performance. MPC–hippocampal bias reliability maps (t-scores) were then converted to normalized z-scores and projected to three surface layers using nearest-neighbor interpolation. The average across the three surface layers was taken as the final surface bias value. In order to test whether areas responsive to recognition demonstrated different patterns of connectivity to the hippocampus from regions implicated in the semantic content analysis, clusters from those analyses (generated separately from even and odd runs) were used as masks. Bias scores were similar across hemispheres and as such combined for statistical comparisons and visualization. First, consistency across even and odd masks was tested using mixed effects linear regression, and then differences in hippocampal connectivity bias were examined across recognition and semantic clusters. Additionally, we investigated whether person and place MPC regions demonstrated differential connectivity with the hippocampus. To do so, we averaged the normalized hippocampal bias scores from within each subject's person- and place-selective clusters. We then submitted these scores to a mixed effects linear regression to compare with a chance level of zero bias across subjects.
Resting-state connectivity analysis
Resting-state functional connectivity was measured using the CONN toolbox 21.a (Whitfield-Gabrieli and Nieto-Castanon, 2012) in MATLAB 9.11 and using SPM12 (Friston et al., 2007) functions. The raw timeseries from the first four resting-state sessions (from experimental Sessions 21 and 22) for each subject were included. Basic preprocessing performed within the CONN toolbox included motion realignment and unwarping, outlier detection for later scrubbing, and direct rigid registration to the 1.0 mm T1 structure. Smoothing (3 mm) was applied to make file sizes more manageable for analysis. Next, denoising was performed by linear regression, removing noise components identified from white matter and CSF timeseries, as well as noise stemming from estimated subject motion. Scrubbing was also applied to limit the impact of outlier TRs, identified during preprocessing. Lastly, recognition and semantic content cluster masks from across hemispheres were used as seed regions. Seed-to-voxel correlation maps were computed as the Fisher-transformed bivariate correlation coefficients between the mean ROI timeseries and each individual voxel timeseries. Finally, resting-state correlation bias scores were calculated in a similar manner as the task-based correlation bias scores above, whereby recognition and semantic cluster correlations were averaged separately for all anterior and posterior hippocampal voxels and then subtracted to create a hippocampal bias score (collapsed across hemisphere). Cluster-extracted hippocampal bias scores were entered into a linear mixed effects regression to evaluate the relationship between clusters and hippocampal connectivity during rest.
Data availability
All data reported here are derived from the open-source NSD (https://naturalscenesdataset.org; Allen et al., 2022).
Results
To examine the functional contributions of human MPC to recognition memory, we leveraged the NSD (Allen et al., 2022), which comprises high-resolution fMRI data collected during an extended recognition memory paradigm as well as resting state (see Materials and Methods). In the recognition memory task, subjects were serially presented with images of natural scenes (photographs) and asked to perform a simple recognition decision of “old” (seen image before) or “new” (never seen image before; Fig. 1c). Importantly, this paradigm was performed across multiple scanning sessions, from which the first 12 sessions (totaling 9,000 trials per subject) were used for all reported task analyses (Fig. 1d; see Materials and Methods). Together, the NSD provides a unique opportunity to examine the functional organization of human MPC during recognition memory decisions, with both high spatial resolution and robust statistical power.
Functional neuroanatomy of item recognition and semantic processing in the MPC
Across sessions, subjects performed well on the recognition memory task (Fig. 1d; mean accuracy = 80.34%; 95% CI = [76.55%, 84.21%]). We examined BOLD fMRI responses within the MPC while subjects performed the stimulus-recognition task to identify regions involved in successful recognition decisions and/or the representation of stimulus content during these decisions. MPC regions responsive to recognition decisions were identified as those where univariate BOLD responses were greater for correct recognition of old stimuli (hits) than for correct identification of new stimuli (correct rejections; Fig. 1c,d,f). However, semantic regions within MPC were identified as those where BOLD activity patterns during recognition were correlated with a semantic model applied to each image (Fig. 1e,f; see Materials and Methods).
Analyses revealed distinct recognition and semantic responsive regions within the MPC for each subject. Strikingly, for all subjects, we found that these functional regions demonstrated minimal spatial overlap. However, the exact spatial configuration and extent of recognition and semantic regions within the MPC varied across individuals (Fig. 2a). For example, while recognition responses (hits > CR) for most subjects were observed in the dorsal PCC/RSC (dPCC/dRSC) and ventral precuneus (vPrC), the spatial extent of these clusters differed across individuals. In all subjects, recognition clusters were identified in the dRSC, while in five of the eight subjects, reliable recognition responses were also identified in the dPCC. As can been seen in Figure 2a, whether this dorsal recognition cluster was primarily located in the dPCC or dRSC varied from subject to subject, highlighting these subtle, but important, individual differences in activation patterns. Additionally, in all but two of the subjects, a reliable recognition cluster was found in the middle of the PCC, around the splenial sulcus (Willbrand et al., 2022, 2023). However, the exact location of this middle PCC recognition cluster was highly variable across individuals. Interestingly, when collapsing across individuals to perform a standard group analysis, only the dRSC and vPrC clusters survived thresholding (Fig. 2b). These two group-level recognition clusters directly overlap with regions associated with the cognitive-control and parietal memory networks (CCN and PMN) within the MPC as identified via previous resting-state functional connectivity analyses (Vincent et al., 2008; Yeo et al., 2011; Shirer et al., 2012; Gilmore et al., 2015). Overall, these group data replicate prior observations of consistent MPC subregional engagement during successful recognition decisions, specifically coactivation of distinct dRSC/PCC and vPrC clusters. However, our ability to robustly quantify individual subject responses suggests that such group maps greatly underestimate the degree of individual variability in MPC recognition responses.
Within the MPC, responses associated with stimulus semantic content also demonstrated variability in spatial configuration and extent across individuals (Fig. 2a). Importantly, unlike the univariate contrast used to identify recognition regions, semantic regions were identified based on displaying activity patterns during recognition that were correlated with a semantic model applied to each image (see Materials and Methods). In all subjects, responses associated with stimulus semantic content were observed within the parieto-occipital sulcus. For four of the eight subjects, this cluster also extended into the gyrus of the ventral PCC. For all but one subject, there was an additional semantic cluster in the dorsal precuneus (dPrC). Additionally, in some subjects, a semantic cluster was identified around the splenial sulcus, anterior to the splenial recognition region noted above. Furthermore, in five of the eight subjects, a semantic cluster was observed around the marginal ramus of the cingulate sulcus. Interestingly, semantic group maps showed widespread organization scattered through much of the MPC (Fig. 2b). This group result indicates that while within-individual evidence for semantic content was below our clustering threshold, across subjects this subtle relationship was consistently present. Group maps of semantic responses overlap with those commonly observed for the canonical DMN via resting-state and task data (Yeo et al., 2011; Shirer et al., 2012). However, depending on the parcellation scheme used, these semantic ROIs may also be considered to overlap with the contextual association network (Doucet et al., 2011; Gordon et al., 2017a) or distinct subnetworks of the DMN (e.g., DMN-A vs DMN-B; Braga and Buckner, 2017; Braga et al., 2019; DiNicola et al., 2020). Additionally, semantic clusters generated here are similar to previous results from studies investigating not only semantic content (Binder, 2016) but also importantly autobiographical retrieval (Addis et al., 2007; Spreng et al., 2009; Andrews-Hanna et al., 2014; Renoult et al., 2019). As such, our semantic group-level results reflect a network of MPC regions that have been previously identified during recollection tasks involving mnemonic representation.
Given the striking variability of MPC response patterns between individuals, we sought to leverage the large number of trials in the dataset to ensure robust and reliable findings. To do so, we used a cross-validation approach, where task trial data was split in half (even/odd) and the analyses performed within each data partition. Consistency across the two data partitions was evaluated by mixed effects linear regression, revealing there was no main effect of data partition (odd/even runs) on analysis results (beta = 0.178; t(7) = 1.817; 95% CI = [−0.235, 0.592]; p = 0.112). Post hoc tests revealed that this lack of difference between data partitions was separately true for the recognition (t(7) = 0.680; mean = 0.086; 95% CI = [−0.213, 0.385]; p = 0.519) and semantic analyses (t(7) = 1.987; mean = 0.271; 95% CI = [−0.052, 0.593]; p = 0.087). The consistency of responses across separate runs of the task suggests that observed patterns of recognition decision responses and semantic content similarity scores were reliable and well founded within individuals. Having established analysis results that were consistent across runs (visualized in Fig. 2c), the following analyses collapsed across partitions when appropriate.
Given the striking interdigitation of recognition and semantic clusters within MPC, we next sought to quantify the degree of spatial dissociation and functional specificity of these clusters. To do so, we compared the mean normalized responses of recognition and semantic clusters, within subjects, for both “recognition decision” and “semantic similarity” analyses. Specifically, we quantified the degree to which recognition clusters displayed recognition or semantic activity in the held-out data-fold and the degree to which semantic clusters displayed recognition or semantic activity in the held-out data-fold. We then collapsed these scores across the two data-folds. As expected, recognition responses were significantly greater in regions masked by recognition clusters than by semantic clusters (t(7) = 10.449; mean = 3.962; 95% CI = [3.066, 4.859]; p < 0.001; Fig. 2c). The opposite relationship was observed for responses extracted by the semantic analysis, whereby values masked within semantic clusters were significantly higher than those masked within recognition clusters (t(7) = −11.83; mean = −2.299; 95% CI = [−2.758, −1.839]; p < 0.001; Fig. 2c). As follows, there was a significant interaction between cluster type and analysis type (t(28) = 17.460; beta = 6.261; 95% CI = [5.527, 6.996]; p < 0.001; Fig. 2c), indicating that response values were greater when the cluster type matched the score-generating type. There was no main effect of response magnitudes being greater in the recognition analysis versus the semantic analysis (t(29) = 0.952; beta = 0.578; 95% CI = [−0.664, 1.820]; p = 0.349) nor was there any main effect of the magnitude of responses being overall greater across both analyses in recognition versus semantic clusters (t(29) = −1.370; beta = −0.8312; 95% CI = [−2.074, 0.410]; p = 0.181). In summary, the combination of cross-run reliability and across-analysis dissociation demonstrates that subregions of the MPC were biased toward either recognition or semantic responses.
Next, we investigated whether responses in recognition and semantic clusters would differentially relate to memory task behavior. First, we compared the correlation between BOLD responses and reaction time (RT) within recognition and semantic clusters. Across subjects, we found that correlations with RT were significantly different between recognition and semantic clusters (t(7) = 4.85; beta = 2.67; 95% CI = [1.59, 3.75]; p = 0.002). Subsequent analyses revealed that there was no reliable relationship between responses and RT in recognition clusters (t(7) = −0.503; mean = −0.395; 95% CI = [−2.252, 1.462]; p = 0.631), but that there was a positive correlation within semantic clusters (t(7) = 6.237; mean = 2.275; 95% CI = [1.412, 3.137]; p < 0.001). Additionally, we considered the relationship between BOLD responses and trial type (easy-old, hard-old, and correct-rejections) to evaluate whether the type of memory decision involved in retrieval modulated responses in recognition or semantic regions. Here, we again observed dissociable responses in the recognition and semantic clusters across trial types. Specifically, we found that while responses in recognition clusters increased with trial type (easy-hit > hard-hit > correct-rejection; t(22) = −18.18; beta = −4.56; 95% CI = [−5.05,−4.07]; p < 0.001), no such relationship was observed in semantic clusters (t(22) = 0.469; beta = 0.136; 95% CI = [−0.43, 0.70]; p = 0.644). Interestingly, responses in semantic clusters were fit by an inverted u-shaped second-order polynomial (t(21) = −5.159; beta = −1.759; 95% CI = [−2.427, −1.091]; p < 0.001), where responses were greatest to hard-old trials, followed by correct-rejection and easy-old trials. These results further serve to dissociate response patterns in recognition and semantic MPC regions.
To confirm the spatial dissociation between recognition and semantic regions in the MPC, we compared the percentage of surface vertices that were uniquely found within recognition, semantic, or both cluster types. As can be seen in the individual and group surface maps (Fig. 2a,b), there was a striking lack of overlap between the two sets of clusters (Fig. 2d). On average within subjects, 19% of MPC vertices were observed to have above threshold recognition responses, while 19% reliably demonstrated semantic content, with only 3% of vertices being implicated in both. Therefore, on average 59% of MPC vertices demonstrated no reliable activity for either response type. Indeed, even at the group level, where a greater number of vertices were associated with semantic content than for any single individual (57%), the overlap between recognition and semantic vertices was still minimal (6%). While cluster evaluation can be influenced by statistical thresholding parameters, particularly searchlight cluster size, recognition and semantic cluster organization and response strength were consistent across data-folds. Furthermore, the use of TFCE (Smith and Nichols, 2009; Winkler et al., 2014) produced a qualitatively similar patterns of results to those reported above (recognition = 16%; semantic = 29%; overlap = 8%).
While thresholding methods allowed for the identification of MPC clusters that reliably demonstrated an association with either recognition decisions or semantic processes, these analyses did not describe the nature of the relationship between recognition and semantic responses. We next sought to evaluate this relationship in more detail by performing two further analyses. First, we examined the correlation of the across-session recognition and semantic responses (z-scores) across all MPC vertices for each subject. Across subjects, we observed a significant negative correlation between recognition and semantic responses (t(7) = −3.098; mean r = −0.224; 95% CI = [−0.396, −0.053]; p = 0.017). Next, to gain additional insight into the relationship between memory decisions and semantic representation, we separately performed the semantic correlation analysis using only hit or correct-rejection trials. Importantly, we found a high degree of overlap between clusters identified using each trial type, with an average of 93% (SE = 3%) of vertices that were identified from the hit trials overlapping with vertices identified from the correct-rejection trials. While there was strong overlap in MPC clusters observed across both hit and correct-rejection analyses, we did observe differences in the strength of correlations across the two memory behaviors. Semantic correlations were reliably weaker for hits than correct-rejections (t(7) = −2.64; beta = −0.50; 95% CI = [−0.87,−0.13]; p = 0.033; across-session mean t-score correct-rejection = 4.406; hit = 3.906). Together, these results suggest that a complex interaction between recognition and semantic responses exists in the MPC. Even though many MPC regions demonstrate a bias toward supporting recognition decisions or semantic representations, there may be variable degrees of processing overlap between these processes that are neither necessarily exclusive nor competitive.
Hippocampal subregion connectivity bias within functional regions of the MPC
Results from our previous analyses suggested a robust functional distinction between MPC regions that are responsive to specific recognition decisions from those that are responsive to the semantic content of recognition stimuli. This dichotomy can be framed as reflecting specific fine-grained details (recognition) compared with more broad conceptual information (semantics) relevant to memory behavior. Strikingly, a similar spectrum of functional specialization has been proposed to exist along the long axis of the hippocampus. Specifically, the posterior hippocampus is thought to support fine-grained mnemonic details, while the anterior hippocampus is proposed to support a more broad “gist” representation (Robin and Moscovitch, 2017). Consistent with this, previous work using functional connectivity measures of resting-state data has suggested that vPCC regions that are part of the DMN (as well as the contextual association network) are coupled with the anterior hippocampus, while dPCC regions that are part of the CCN/PMN are more strongly coupled with the posterior hippocampus (Zheng et al., 2021). Given these observations, we predicted that a similar profile of hippocampal connectivity bias would be observed for our functionally defined MPC recognition (greater posterior hippocampal coupling) and semantic (greater anterior hippocampal coupling) clusters, within individuals.
To compare the hippocampal connectivity profiles of functionally defined MPC regions, we independently calculated the average correlation of MPC recognition and semantic clusters to the anterior and posterior aspects of the hippocampus during both task and rest (see Materials and Methods). We subtracted the difference between hippocampal correlations (anterior–posterior) for each cluster to obtain a connectivity bias score, where negative values indicated stronger correlations to the posterior hippocampus and positive values indicated stronger correlations to the anterior hippocampus (Fig. 3a–c). Using mixed effects linear modeling, we first evaluated whether there was any effect of data-fold (even/odd), hippocampal hemisphere (left/right), MPC hemisphere (left/right), or any interactions therein, on the connectivity bias measure during recognition task performance. As this analysis showed no significant effects of these variables on hippocampal connectivity bias (all p > 0.5), we collapsed across data-fold and hemisphere to investigate the relationship between hippocampal bias and MPC clusters (cluster type: recognition/semantic) during task performance. Interestingly, there was a significant main effect of cluster type (t(7) = 6.289; beta = 3.577; 95% CI = [2.232, 4.922]; p < 0.001; Fig. 3d), with recognition clusters consistently demonstrating a stronger connectivity bias toward the posterior hippocampus than semantic clusters. Post hoc testing revealed that recognition cluster connectivity was significantly biased toward the posterior hippocampus (t(7) = −19.968; mean = −4.374; 95% CI = [−4.892, −3.856]; p < 0.001), while semantic clusters demonstrated a dissociable pattern of connectivity that was characterized by equivalent correlation strength across the length of the hippocampus (t(7) = −1.301; mean = −0.797; 95% CI = [−2.246, 0.652]; p = 0.235). To further examine this relationship, we performed the reciprocal analysis of quantifying the connectivity bias of hippocampal voxels, across its longitudinal axis, with recognition versus semantic regions (Zheng et al., 2021). In doing so, we observed that for all subjects during task and rest, the posterior hippocampus demonstrated greater connectivity toward recognition regions, with the anterior hippocampus reflecting a gradual shift in connectivity bias toward semantic regions (task: t(7) = −6.573, beta = −0.022, 95% CI = [−0.30,−0.014], p < 0.001; rest: t(7) = −9.135, beta = −0.084, 95% CI = [−0.105, −0.062], p < 0.001).
Functional connectivity between the MPC and hippocampal subregions during task and rest. a, Example beta-series of single voxels selected from a recognition cluster (red) and semantic cluster (blue) within the MPC (left), as well as anterior (green) and posterior (purple) hippocampus (right) during one task session. b, Scatterplots show the correlation of beta values (one session) from either the recognition voxel (top) or semantic voxel (bottom), with both the anterior and posterior hippocampal voxels. Each point represents the percent signal change from a representative MPC voxel and representative anterior (green) or posterior (purple) hippocampal voxels on each recognition trial. Lines represent the linear relationship across all trials from one session between the representative MPC voxel and hippocampal voxels. c, Cortical surface map showing the across-session MPC–hippocampus connectivity bias values (t-scores) for S1. The surface color map indicates greater connectivity from the MPC to the anterior (green) or posterior (pink) hippocampus. Outlines indicate the subject's recognition (red) and semantic (blue) clusters used as ROIs. d, The average normalized (z-score) connectivity bias for voxels in recognition (red) and semantic (blue) ROIs from task (left) and rest (right). Voxels within recognition clusters demonstrated a significant connectivity bias to the posterior hippocampus, while no such relationship was observed for semantic cluster voxels.
Next, we performed the same hippocampal bias analysis using resting-state data (see Materials and Methods), with individual task identified MPC clusters serving as ROIs. Recognition and semantic ROIs demonstrated statistically similar MPC–hippocampal connectivity whether they were generated from the even or odd data-folds (t(7) = 0.814; mean = 0.027; 95% CI = [−0.0513, 0.105]; p = 0.442). Similar to task data, we observed a main effect of ROI type (t(7) = 9.590; beta = 2.221; 95% CI = [1.673, 2.768]; p < 0.001; Fig. 3d), with recognition ROIs consistently demonstrating a stronger posterior hippocampal bias than semantic ROIs. Post hoc testing revealed that for resting-state data, recognition ROIs demonstrated a significant connectivity bias to the posterior hippocampus (t(7) = −7.820; mean = −2.329; 95% CI = [−3.033, −1.624]; p < 0.001), while semantic clusters demonstrated connectivity biased to a broader length of the hippocampus (t(7) = −0.745; mean = −0.108; 95% CI = [−0.451, 0.235]; p = 0.481). Together, these connectivity results further support a striking dissociation between recognition and semantic regions of the MPC, one which extends into distinct profiles of connectivity with the hippocampus.
Categorical selectivity to people and places in the MPC
In identifying a putative functional distinction between MPC regions supporting stimulus-recognition decisions from those supporting stimulus semantic content, it is worth considering the features of this semantic representation, particularly as they relate to common elements of episodic memories. Recently, it has been observed that the MPC demonstrates categorical selectivity for people and place stimuli, common features of episodic memory, during perception and memory retrieval (Visconti di Oleggio Castello et al., 2017; Silson et al., 2019; Afzalian and Rajimehr, 2021; Steel et al., 2021, 2023; Srokova et al., 2022). This observation is surprising, as historically the MPC has not been implicated in specific sensory processes (Margulies et al., 2016), consistent with its general lack of direct connectivity with primary sensory regions (Parvizi et al., 2006). However, if the semantic MPC regions identified in the analyses above are indeed involved in the broad representation of recognition stimulus content, then putative person- and place-selective regions should coincide within MPC semantic clusters. In addition, our findings would suggest that such an overlap is sensitive to performing mnemonic tasks.
The NSD provides a unique opportunity to examine both of these predictions, as it also includes a standardized visual category localizer task (fLoc) for each subject (see Materials and Methods). This dataset serves as a control for the main recognition task, by including unfamiliar, category-specific cropped grayscale images (i.e., not natural images) presented during a nonrecognition task (fixation with oddball “background-only” stimulus detection). It secondarily serves as a benchmark for identifying visual category selectivity within subjects. Therefore, to examine these questions we (1) mapped people- and place-selective regions across the cortex in each subject using the standardized visual category localizer task (fLoc), (2) compared these functional maps with those constructed using only people and place stimuli from the NSD recognition task, and (3) specifically compared the degree of overlap within the MPC for people/place maps generated via the fLoc and NSD tasks, with identified semantic clusters.
First, we mapped people- and place-selective regions across the cortex in each subject using the fLoc task, where subjects were presented 10 visual categories, for which two categories were combined to define person (faces: adult and children) and place (houses and corridors) conditions (Fig. 4a,c). Consistent with a large literature, person-selective clusters were robustly and reliably observed in the lateral occipital, fusiform gyrus, and anterior temporal lobe face areas (Tsao et al., 2008; Rajimehr et al., 2009; Grill-Spector et al., 2017). Also consistent with prior work, place-selective clusters were observed in the lateral occipital, parahippocampal gyrus, and parieto-occipital place areas (Bainbridge et al., 2021; Steel et al., 2021). Similarly, robust evidence for person- or place-selective responses was not observed in the MPC, where only sporadic, limited person and place selectively responsive clusters were observed outside of the parieto-occipital place area (i.e., medial place area; Steel et al., 2021). This finding is predominately due to much of the MPC displaying consistent univariate deactivation to task stimuli. However, growing evidence suggests deactivation profiles in the MPC may convey broad visual category (Silson et al., 2019) and retinotopic tuning (Szinte and Knapen, 2020). In this regard, while fLoc results did not survive thresholding within subjects, we did observe a general trend that deactivations in the dorsal MPC for faces were smaller in magnitude than for other categories across subjects. Using this established visual category localizer task, we were able to reliably map expected person- or place-selective cortices; however, we observed little evidence for person or place selectivity within the MPC during perception.
Categorical selectivity within the MPC. a, Top: stimuli from the category fLoc task for either the person (adult and child faces) or place (houses or corridors) categories. Bottom: examples of stimuli that were included (left column) or excluded (right column) as person or place stimuli in the NSD task categorical selectivity analysis. b, Regions demonstrating selective responses during the NSD task toward people versus places visualized on the medial inflated surface anatomy for each subject. c, Ventral surface visualization from one subject showing the spatial overlap of person (white outline) and place (black outline) clusters from the category localizer and person (green) and place (orange) clusters from the NSD task. d, Person (white)- and place (black)-selective clusters from the category localizer demonstrated similar selectivity to person (negative values) and place (positive values) during the NSD task. The categorically biased responses were similar across even and odd runs of the NSD task (x-axis). e, Person and place clusters from the NSD task were used as masks for the recognition and semantic analysis. While there was no significant recognition response found within person or place clusters overall, significant semantic content responses were observed in both cluster types. Throughout, green is used to indicate person clusters and orange indicates place clusters identified from the NSD recognition task. NSD task clusters were thresholded at z > 3.3 with at least 20 contiguous vertices. Category localizer clusters were thresholded at t > 3.3 and at least 20 contiguous vertices.
Next, we performed a similar analysis using selected person and place stimuli from the NSD recognition task in order to benchmark the identification of person/place selectivity against maps generated from the fLoc task above (see Materials and Methods). With this goal in mind, we compared responses to person/place stimuli during the recognition task within person/place clusters identified via the fLoc task across the whole brain. As expected, activity within person and place fLoc clusters demonstrated categorical selectivity during the recognition task (paired t test: t(15), 24.208; mean = 7.361; 95% CI = [6.713, 8.009]; p < 0.001; Fig. 4d). Post hoc analyses demonstrated that both person and place clusters generated from the fLoc task respectively showed high person (t(7) = 18.531; mean = 3.786; 95% CI = [3.302, 4.269]; p < 0.001) and high place (t(7) = −11.900; mean = −3.575; 95% CI = [−4.286, −2.865]; p < 0.001) selectivity during the recognition task. These results confirmed a strong consistency in cortical regions identified as person- or place-selective via a well-controlled localizer task or person/place trials of the NSD recognition task.
We next investigated whether categorically selective responses within the MPC were equivalent across the fLoc and recognition tasks. As the NSD task has many more trials than the fLoc task, and therefore greater statistical power, we examined the strength of categorical selectivity for each individual NSD session, where trial number is comparable to the fLoc (see Methods and Materials). Across subjects, this analysis showed that on average MPC responses were consistently greater for trial-matched NSD sessions than the fLoc for both person (t(7) = 5.974; beta = 1.599; 95% CI = [1.043; 2.155], p < 0.001) and place (t(4) = 4.642; beta = 1.055; 95% CI = [0.566, 1.544]; p = 0.010) stimuli. In contrast, however, in the parieto-occipital cortex, a region consistently associated with selective responses during place perception (Silson et al., 2019; Steel et al., 2021), the inverse relationship was observed with stronger responses to the fLoc compared with the NSD task for place stimuli (t(7) = 2.620; beta = 0.669; 95% CI = [0.065, 1.272]; p = 0.034). These observations support the efficacy of identifying person-/place-selective regions by the NSD recognition task and also suggest that selectivity to these categories within the MPC may be sensitive to mnemonic processes (Silson et al., 2019; Afzalian and Rajimehr, 2021; Deen and Freiwald, 2022).
In light of this task difference, we next sought to map the functional organization of people-/place-selective regions within the MPC across individuals. While the location of selective regions varied greatly between subjects, certain patterns did emerge (contrast maps are shown in Fig. 4b). For example, while all eight subjects exhibited at least one person-selective patch within the MPC in both hemispheres, the location, number, and extent of person-selective regions varied from subject to subject. Across individuals, person-selective clusters differed in whether they were found in the parieto-occipital sulcus, splenial sulcus, or near the marginal ramus. Regarding places, all but one subject exhibited a selective region in the inferior parieto-occipital sulcus. For most subjects, this cluster did not extend into the neighboring gyrus of the vPCC. Together, we observed distinct person- and place-selective regions within the MPC, consistent with prior observations (Silson et al., 2019); however, our precision data suggested that these maps showed anatomical variability across subjects, much like identified semantic clusters.
To quantify functional overlap between the person- and place-selective regions of the MPC with the recognition and semantic clusters identified above, we conducted a mixed effects linear regression, comparing recognition and semantic content responses within identified person and place clusters of the MPC (Fig. 4e). Across both person and place clusters, we found a main effect such that average responses were significantly higher for the semantic analysis than the recognition contrast (t(47) = 17.424; beta = 4.049; 95% CI = [3.573, 4.524]; p < 0.001). Post hoc tests revealed that overall the average recognition responses within person and place clusters did not significantly differ from zero (t(7) = 0.756; mean = 0.252; 95% CI = [−0.537, 1.042]; p = 0.474; Fig. 4e), nor did recognition responses significantly differ between person and place clusters (t(13) = 1.138; mean = 0.341; 95% CI = [−0.307, 0.990]; p = 0.276). Conversely, semantic responses from the same person and place clusters were significantly greater than zero (t(7) = 32.12; mean = 4.188; 95% CI = [3.880, 4.497]; p < 0.001) and were similar between person and place clusters (t(13) = −0.770; mean = −0.123; 95% CI = [−0.468, 0.222]; p = 0.455).
Next, we looked at the degree of overlap between categorically selective clusters in the MPC (collapsing across both person- and place-selective clusters) with recognition and semantic clusters (recognition task data). After averaging across data-fold and hemisphere within subjects who had at least one person- or place-selective MPC cluster (person, n = 8; place, n = 7), we found that 75% of the person-selective cluster vertices and 69% of the place-selective cluster vertices overlapped with semantic clusters, while only 1% of person-selective and 3% of place-selective cluster vertices overlapped with recognition clusters. A remaining 19% of person-selective and 24% of place-selective cluster voxels fell outside of either the recognition or semantic clusters, while the remaining 5% of person- and 4% of place-selective vertices overlapped with clusters that demonstrated both recognition and semantic responsivity. While thresholding can impact cluster identification, the organization of person and place responsive regions in the MPC was similar when cluster-free threshold enhancement was applied (e.g., on average across subjects 79% of both person and place cluster vertices overlapped with semantic clusters). Together, these data further support the view that identified semantic clusters within the MPC contain subregions displaying selectivity to person or place stimuli, specifically when making mnemonic-based judgments.
We observed significant overlap between semantic and person clusters, primarily located near the splenial sulcus, as well as semantic and place clusters, primarily located near the parieto-occipital sulcus and vPCC/vRSC. This observation converges with studies of fMRI resting-state functional connectivity which often treat these anatomical regions of the MPC as part of a single DMN (Yeo et al., 2011). However, studies that have allowed for more precise delineation of functional neuroanatomy in functional networks have also considered the person and place MPC regions as members of two closely linked but separate networks (Yeo et al., 2011; Braga and Buckner, 2017; Gordon et al., 2018). We therefore also analyzed whether person and place MPC regions could be dissociated based on functional connectivity to the temporal lobe (i.e., a characteristic that divides the broader recognition and semantic regions). We found that MPC person areas demonstrated stronger connectivity with the FFA than the PPA, and place regions demonstrated the inverse relationship (t(13) = 15.50; beta = 23.50; 95% CI = [20.52, 26.47]; p < 0.001). However, there were no differences between MPC person and place areas in terms of connectivity bias between the anterior and posterior hippocampi (t(6) = 1.038; beta = 0.413; 95% CI = [−0.366, 1.191]; p = 0.339). In one respect, we replicate previous work demonstrating dimensions by which person and place MPC regions may be dissociated (categorically selective activity to people vs places and related functional connectivity with the ventral temporal cortex). However, we also observed that these MPC regions demonstrate similar recognition decision responses, connectivity with the hippocampus, and spatial overlap with semantic MPC clusters. Together, these results suggest that person and place MPC regions may operate as parts of an overarching semantic network.
Discussion
We analyzed a high-resolution open fMRI dataset (NSD; Allen et al., 2022) collected while subjects performed thousands of stimulus-recognition trials to characterize the functional contributions of MPC to recognition memory. Within individuals, we were able to identify two interleaved, though largely nonoverlapping, subnetworks: one associated with recognition memory decisions and the other that reflected the semantic content of mnemonic stimuli. These two subnetworks were further differentiated by their profile of functional connectivity with the hippocampus and responsivity to person and place stimuli. Recognition regions demonstrated biased connectivity with the posterior hippocampus and no categorical selectivity to people or places. Conversely, semantic regions were more strongly correlated with the anterior hippocampus than recognition regions and did demonstrate categorically selective responses to people and places. Importantly, we observed robust idiosyncrasies in the spatial organization of recognition and semantic regions within individuals that differed substantially from group maps. When taking these individual differences into account, our results demonstrate how MPC dissociations previously identified by distinct forms of task, network, and stimulus selectivity reflect analysis specific facets of a common organizational principle.
Within individuals, we observed separate, though closely interdigitated, recognition and semantic responsive MPC regions. Recognition responses were consistently found in the dRSC and dPCC, as well as the vPrC and near the splenial sulcus. Meanwhile, semantic regions were observed within the parieto-occipital sulcus, vPCC, mid-PCC, and dPrC. At the group level, recognition regions were limited to the dRSC and vPrC, mirroring previous results from item-recognition studies (Kim, 2013; Gilmore et al., 2015). Group-level semantic clusters were observed from the parieto-occipital sulcus and vPCC to the dPrC, reflecting similar results that have been found in analyses of semantic representation (Binder et al., 2009; Binder and Desai, 2011; Lee and Chen, 2022a) and episodic recollection (Andrews-Hanna et al., 2014; Bird et al., 2015; Kim, 2021). Additionally, this group-level functional neuroanatomy mimics dissociations reported in studies directly comparing MPC responses to item recognition with autobiographical retrieval (McDermott et al., 2009; H-Y. Chen et al., 2017), as well as studies comparing content learned in a laboratory setting with that in lived experience (Elman et al., 2012, 2013). Furthermore, similar dissociations have been observed in studies examining episodic recollection versus familiarity (Henson et al., 1999; Yonelinas et al., 2005), which have led to the proposal that patterns of activity in the MPC are related to the level of cognitive control versus mnemonic representation involved during retrieval (Kim, 2021). Supporting this view is the observation that recognition clusters overlap with the frontoparietal control or parietal memory networks (Kwon et al., 2023), associated with cognitive control and familiarity processing (Dosenbach et al., 2008; Gilmore et al., 2015), while semantic clusters overlap with the DMN, classically associated with autobiographical memory and internally directed cognition (Buckner et al., 2008; Menon, 2023). Indeed, converging evidence from human and animal studies suggests that MPC recognition regions may serve a broader executive function, not limited to mnemonic processing (for review, see Foster et al., 2023). For example, activity in the dPCC/dRSC and vPrC have consistently been linked with value and decision-outcome processing (Kable and Glimcher, 2007; Bartra et al., 2013; Clithero and Rangel, 2014; Oldham et al., 2018; Wolff et al., 2020), environment updating processes (Visalli et al., 2019; Wolff and Brechmann, 2023), event segmentation (Lee and Chen, 2022b), decision-risk evaluation (Kolling et al., 2014), goal congruency (Frömer et al., 2019), explore versus exploit decisions (Barack et al., 2017), saliency processing (Heilbronner et al., 2011), memory-guided attention (Rosen et al., 2016, 2018), and executive control (Leech et al., 2011; Aponik-Gremillion et al., 2022). These dPCC/dRSC results associating the subregion with executive or decision-making processes are in stark contrast to neuroimaging studies implicating the vPCC, which largely link this subregion to the recollection of experiences (Benoit and Schacter, 2015; Bird et al., 2015; Gilmore et al., 2021). Our results converge with growing evidence that a functional dissociation exists within the MPC during episodic recollection and recognition, which distinguishes regions involved in executive control during memory-based decisions from regions involved in the retrieval and representation of mnemonic content.
What factors may help account for this dissociation of functional regions observed in MPC? Interestingly, we found within subjects that recognition and semantic clusters demonstrated distinct patterns of functional connectivity with the hippocampus during task performance and rest. Specifically, we observed that recognition clusters exhibited a strong connectivity bias toward the posterior versus anterior hippocampus. Conversely, semantic clusters were observed to have a broader connectivity bias between the anterior and posterior aspects of the hippocampus and consistently weaker posterior connectivity strength than recognition clusters. Given that such task/rest coupling often conveys putative functional association (Biswal et al., 1995; Dosenbach et al., 2007; Foster et al., 2015, 2016), recent progress in revealing a hierarchy of organization throughout the longitudinal axis of the hippocampus can serve to help better understand the nature of our dissociations within the MPC. For example, during memory retrieval, the anterior hippocampus is associated with broad, gist processing involved in autobiographical retrieval, while the posterior hippocampus is linked to more specific, detailed representations in support of goal-directed processing (Robin and Moscovitch, 2017; Sekeres et al., 2018; Zheng et al., 2021). More generally, the hippocampus is thought to be characterized by a gradient of representational granularity that shifts from broad to more fine-tuned representation along the anterior to posterior axis, respectively (Collin et al., 2015; Brunec et al., 2018; Raut et al., 2020). Thus, while recognition and semantic regions of the MPC may primarily perform dissociable processes in parallel, the degree of observed activity within each subregion will be dependent on the balance of fine-grained memory-related control processes versus the broad reinstatement of episodic details required by tasks.
The recognition and semantic subnetworks of the MPC are further differentiated by their selective responses to categorical stimuli. Using the naturalistic stimuli from the NSD, we found categorically selective responses in the MPC when individuals were making memory decisions. Importantly, our results confirmed that categorically selective responses were predominantly located within semantic and not recognition clusters, suggesting that identified person/place regions are part of a larger network involved in broader semantic or contextual association representation, rather than solely in the specialized processing of perceptual attributes of stimuli (Bar et al., 2008a,b; Bonner and Epstein, 2021). Our findings also build on previous reports of categorically selective responses reported in the MPC during recognition of experimenter-controlled people and place stimuli and during recall of highly familiarized memories of famous or personally relevant people and places (Silson et al., 2019; Woolnough et al., 2020; Gilmore et al., 2021; Hill et al., 2021; Deen and Freiwald, 2022). Indeed, previous work has found that categorically selective responses in the MPC during retrieval are modulated by the degree of familiarity one has with a given stimulus (Visconti di Oleggio Castello et al., 2017; Silson et al., 2019; Afzalian and Rajimehr, 2021). We found that, outside of the parieto-occipital sulcus, categorically sensitive responses were greater during the recognition task than during a passive perceptual localizer, adding to the evidence that the responses in the region are mnemonically sensitive. Additionally, these data suggest that features of particular relevance to episodic memory, such as the people and places that make up our social experiences, may be more represented over other feature domains (Kidder et al., 2022), reflecting a hippocampal transformation of ventral stream sensory inputs into the MPC.
To build on our understanding of the role of the MPC in episodic memory and cognition in general, future work should continue to precisely characterize the functional neuroanatomy of the region. For example, while the present work indicates that one set of subregions supports executive processes during episodic memory tasks, further research is needed to determine whether this is the same region that is implicated in broader decision-making processes like subjective-value or decision-outcome evaluation (Kable and Glimcher, 2007; Bartra et al., 2013; Clithero and Rangel, 2014). We also observed that even though MPC regions demonstrate a clear bias toward supporting recognition decisions or semantic representations, these functions need not be exclusive nor competitive. MPC subregions can be anatomically viewed as convergence zones of different executive and mnemonic processes recruited in the common service of supporting episodic remembering (Damasio, 1989). For example, recent work has shown that activity in recognition regions is modulated by semantic content during memory retrieval (Lee et al., 2023). Importantly, while only examining recognition behavior, our identification of both recognition and semantic subregions during this task highlights how retrieval task demands will differentially drive distinct patterns of MPC response, which may fail to fully capture other, more subtle, task relevant activities throughout the MPC (Rosen et al., 2018; Gilmore et al., 2019b; Aponik-Gremillion et al., 2022; Lee et al., 2023).
Results from the current study represent an important step toward more precisely understanding how the MPC contributes to episodic memory and cognition more generally. Using a precision-neuroimaging approach within individuals, we observed that functionally distinct subregions of the MPC supported either memory decisions or memory contents during recognition and that these subregions closely recapitulated the spatial organization of hippocampal connectivity and visual category selectivity within the MPC. This converging evidence suggests that recognition regions support fine-grained evaluation of stimuli in support of mnemonic decision-making. Conversely, semantic regions are involved in representation and retrieval of broad details during episodic retrieval. Therefore, the commonly observed dissociation of episodic retrieval task activity within the MPC is likely driven by the degree to which executive versus mnemonic components are required and that rather than two separate sets of processes, these executive and mnemonic components operate in complex interaction within the MPC as a convergent zone for such neural systems (Foster et al., 2023). These functional attributes may help to better understand the unique contributions of the MPC while more broadly revealing a principle of functional organization common throughout memory systems in the human brain.
Footnotes
This work was supported by NIH Grants F32MH130027 to S.R.K. and R01MH129439 to B.L.F. Collection of the NSD dataset was supported by NSF Grants IIS-1822683 and IIS-1822929.
The authors declare no competing financial interests.
- Correspondence should be addressed to Seth R. Koslov at seth.koslov{at}pennmedicine.upenn.edu or Brett L. Foster at brett.foster{at}pennmedicine.upenn.edu.