Authors
Tianxiao He, Anna Maslarova, Mihály Vöröslakos, Chenyi Li, Yurong Liu, György Buzsáki, Erdem Varol
Abstract
In electrophysiology, precise in vivo localization of recording sites in deep brain structures is crucial for consistent targeting in multi-day recordings and accurate deep brain stimulation. However, current approaches present their own challenges: brain atlas-guided probe insertion may be imprecise due to anatomical variability, CT or MRI scan based localization may lack the spatial resolution for subregion structures, and post hoc histology misses longitudinal information in chronic recordings. Focusing on mouse recordings, we present a learning-based automatic localizer to identify hippocampus sublayers from high-density extracellular recordings. Our framework operates on functional correlations of Local Field Potential between channels and employs latent features in low dimensional manifolds for region identification. Critically, we observe that neural manifolds across sessions and animals share a common geometry, perturbed by simple transformations. To minimize this variability, we align all neural manifolds to a common space using two methods: a linear approach that learns a rotation and translation for each session’s manifold, and a nonlinear neural network approach with supervised contrastive learning to map different subject recordings into a shared embedding. Both methods yield embeddings that minimize cross-subject variability while preserving cross-region variability. This enables us to learn a probabilistic decoder to predict regions of each channel from learned embeddings. We demonstrate our method on mice Neuronexus recordings with 1024 channels over 8 shanks, providing a better spatial coverage than linear probes like Neuropixels. Our predictions are consistent with neuroanatomy of the hippocampus, resembling the layered structures without prior modeling. Within 7 animals’ recordings, it is able to localize 1024 channels with 93% accuracy using only 5% labeled channels, outperforming the baseline model adopted from LOLCAT (Schneider et al. 2023). Across sessions, our model predicts hippocampal regions with 71% accuracy using minimum supervision in the target session. These results suggest that hippocampal regions are distinguishable by their functional correlations with other brain regions, generalizing across subjects and probe designs.