Two Distinct Scene-Processing Networks Connecting Vision and Memory

Abstract

A number of regions in the human brain are known to be involved in processing natural scenes, but the field has lacked a unifying framework for understanding how these different regions are organized and interact. We provide evidence from functional connectivity and meta-analyses for a new organizational principle, in which scene processing relies upon two distinct networks that split the classically defined parahippocampal place area (PPA). The first network of strongly connected regions consists of the occipital place area/transverse occipital sulcus and posterior PPA, which contain retinotopic maps and are not strongly coupled to the hippocampus at rest. The second network consists of the caudal inferior parietal lobule, retrosplenial complex, and anterior PPA, which connect to the hippocampus (especially anterior hippocampus), and are implicated in both visual and nonvisual tasks, including episodic memory and navigation. We propose that these two distinct networks capture the primary functional division among scene-processing regions, between those that process visual features from the current view of a scene and those that connect information from a current scene view with a much broader temporal and spatial context. This new framework for understanding the neural substrates of scene-processing bridges results from many lines of research, and makes specific functional predictions.

Keywords: memory; networks; scene; vision.

Graphical abstract

Figure 1.

Connectivity clustering of cortical parcels. The cortex was first grouped into 172 local parcels (black lines), such that the surface vertices in each parcel had similar connectivity properties. Performing a second-level hierarchical clustering on these parcels identified distributed networks of strongly connected parcels (parcel colors denote their network membership). Scene-related regions of interest (identified using standard scene localizers in a separate group of subjects) are split across two networks, which are largely symmetric across left (top row) and right (bottom row) hemispheres. OPA and posterior PPA overlap with a posterior network (dark blue) that covers all of visual cortex outside the foveal confluence, while cIPL, RSC, and aPPA overlap with an anterior network (magenta) that covers much of the default mode network.

Figure 2.

Connectivity shifts across the network border. a, Using classic multidimensional scaling (MDS), we can visualize the connectivity structure among the eight parcels overlapping with scene-related regions (darker/lighter shading denotes left/right hemisphere). The first MDS dimension shows a parallel transition along both dorsal and ventral paths from parcels overlapping OPA and pPPA to those overlapping cIPL, RSC, and aPPA. b, Connectivity between dorsal parcels and the medial RSC parcel increases markedly near the OPA/cIPL border. b, Ventral parcels also show a shift in network connectivity properties, with increasing connectivity to the most anterior cIPL parcel as we move from pPPA to aPPA. Error bars are 95% confidence intervals across subjects, *p < 0.05, **p < 0.01.

Figure 3.

Connectivity between network parcels and the hippocampus. a, For each parcel in the anterior and posterior scene networks, we computed its resting-state connectivity with the hippocampus, showing a striking increase in hippocampal activity for anterior network parcels overlapping with cIPL, RSC, and aPPA (magenta circles) compared with posterior network parcels (blue circles). b, Along the dorsal network boundary, hippocampal activity first dips slightly and then increases substantially, becoming strongest in the most anterior parcel intersecting cIPL (and is also high in RSC). c, Ventrally along parcels overlapping with PPA, we observe a similar increasing posterior-to-anterior gradient in connectivity. d, Computing the connectivity between each coronal slice of the hippocampus and the two scene networks shows that this increased coupling to the anterior network is present throughout the hippocampus, but is especially pronounced in anterior hippocampus (MNI coordinate y > −21 mm). Error bars are 95% confidence intervals across subjects. *p < 0.05, **p < 0.01.

Figure 4.

Overlap of posterior and anterior scene networks with previous work. a, Two meta-analyses conducted using NeuroSynth identified overlapping but distinct reverse-inference maps corresponding to studies of visual scenes and to studies of higher-level memory and navigation tasks. These maps separate into our two scene networks, with visual scenes activating voxels in the posterior network and memory/navigation tasks activating voxels in the anterior network, as shown on example axial (z = −8) and sagittal (x = −30) slices. False discovery rate < 0.01; cluster size, 80 voxels (640 mm³). b, Voxels having a >50% chance of belonging to a retinotopic map (orange) overlap with much of the posterior scene network, but end near the border of the anterior scene network. Breaking up the contributions of individual regions, we find that the probability mass of the topographic maps falls primarily within the posterior network, with only PHC2 showing a small overlap with the anterior network (probabilistically at the group level).

Figure 5.

Two-network model of scene perception. Our results provide strong evidence for dividing scene-sensitive regions into two separate networks. We argue that OPA and posterior PPA (PHC1/2) process the current visual features of a scene [in concert with other visual areas, such early visual cortex (EVC), and LOC], while cIPL, RSC, and aPPA perform higher-level context and navigation tasks (drawing on long-term memory structures including the hippocampus).