Scene selective medial temporal lobe subregions are recruited for the integration of non-scene stimuli

A hallmark of episodic memory is the ability to flexibly recombine information across episodes to form new associations and guide behaviour. This process, termed associative inference, relies on the hippocampus and surrounding medial temporal lobe (MTL) subregions. We previously found that cross-episode binding is improved when episodes are linked by scenes rather than by faces or objects. Here we tested whether differential recruitment of category-selective MTL subregions underlies these behavioural differences. Participants (n=31) completed study-test phases of the Associative Inference in Memory task, while undergoing high-resolution fMRI scanning at 3 tesla. During the study phase, participants encoded overlapping AB and BC pairs. A and C items were always objects, but the linking B item was either a face or a scene. At test, memory for the direct (AB, BC pairs) and indirect associations (inferred AC pairs) was tested. No face-selective regions were identified within the MTL based on an independent functional localiser scan. Within the MTL, the anterior hippocampal head, a region of interest which included the cornu ammonis 2, 3 and dentate gyrus (CA2-3/DG), anterolateral and posteromedial entorhinal cortex (alERC and pmERC), perirhinal cortex (PRC), and parahippocampal cortex (PHC) were identified as scene-selective (i.e., more active for scenes compared to faces). Although accuracy of the indirect inferences did not differ between associative pairs linked by faces and scenes, MTL cortex subregion recruitment differed across categories. Subregions in the MTL cortex (alERC, pmERC, PRC and PHC), but not the hippocampus (anterior hippocampal head, CA2-3/DG), were recruited to support associative inference for faces during encoding. These findings suggest regions in the MTL cortex identified as scene-selective here may be specialised for integrating disparate elements of episodes into coherent representations, and may be recruited when integration demands during encoding are high (e.g., during associative inference) even for non-scene stimuli.