Densely reconstructed electron microscopy (EM) volumes have ushered in a transformative era in neuroscience, offering unprecedented access to the intricate connectivity between diverse neural subtypes. These volumetric datasets provide a microscopic map of neural circuits critical for understanding brain function at the cellular and subcellular level. However, the challenge remains formidable: classical electron microscopy, despite its spatial precision, does not inherently reveal genetic or molecular markers of individual neurons. Instead, cell typing often relies on morphological cues or connectivity patterns, requiring intensive manual proofreading or sophisticated computational approaches to accurately define cell classes.
Recent advances have leveraged extensive prior work linking morphology to transcriptomic profiles, bridging the gap between structural and molecular taxonomy. Such integrative efforts, including contributions from large consortia, have demonstrated that various features — from the shape of dendritic and axonal arbors to local nuclear and peri-somatic morphology — contain rich information integral to neuron classification. Furthermore, even small local segments of neural processes and multi-scale views of single nuclei have proven surprisingly informative, underscoring the potential of EM data for fine-grained neural stratification.
In this context, the newly developed NEURD framework introduces a powerful suite of automated tools for proofreading and feature extraction tailored to connectomics datasets. NEURD focuses on non-branching dendritic segments as fundamental units, generating a rich, interpretable feature set that enhances the accuracy of cell-type classification. This approach leverages morphological nuances embedded within dendritic graphs, forging a novel path for automated identification workflows that promise to accelerate connectomic research.
To validate the utility of this framework, logistic regression models trained on as few as two synaptic features successfully segmented excitatory from inhibitory neurons across two independent datasets, MICrONS and H01, with impressive accuracy. This cross-dataset consistency highlights the robustness of the extracted features and their generalizability, affirming their biological and computational relevance for basic cell-class separation tasks.
Extending beyond this binary distinction, NEURD was employed in combination with graph convolutional networks (GCNs) to probe finer scales of cell-type identity. GCNs, sophisticated deep learning architectures capable of learning from graph-structured data, were trained on dendritic subgraphs derived from a rich and carefully curated set of hand-labelled neurons within the MICrONS volume. This volume represents one of the most comprehensively annotated EM datasets available, with detailed cell-type labels informed by expert neuroanatomical assessment.
Focusing exclusively on dendrites — which exhibit both high segmentation recall and impressive proofreading precision in this dataset — the GCN was able to embed these complex morphologies into a structured feature space. Strikingly, the majority of cells within the volume, including those outside the labelled training set, occupied expected regions of this embedding, corresponding to their respective cortical laminar positions. This occurred despite no direct use of spatial coordinates in training, underscoring the model’s ability to internalize biologically relevant shape and connectivity features.
Performance on a held-out test cohort demonstrated that the GCN classifier achieved a mean class accuracy of approximately 82%, accurately distinguishing multiple excitatory pyramidal neuron subtypes spanning cortical layers 2 through 6, as well as diverse inhibitory interneuron types like basket, bipolar, Martinotti, and neurogliaform cells. Individual class accuracies were remarkably high for many neuron types, including perfect classification scores for certain basket and Martinotti cells, revealing the nuanced discriminative power encoded within dendritic architectures alone.
Moreover, the study evaluated classification using disconnected dendritic stems — isolated branches connected directly to the soma — to assess how much local morphology contributes to cell identity. Even with these incomplete inputs, the classification performance was only moderately reduced to a mean accuracy of 66%, indicating that local dendritic morphology harbors substantial predictive information about neuronal identity. This finding aligns with and extends previous literature suggesting that fine-scale, local features can serve as reliable discriminants among cell classes.
Importantly, NEURD’s deep learning classifiers provide probabilistic confidence scores through their final softmax layer outputs. These confidence metrics enable researchers to tailor downstream analyses by selecting only high-confidence labels, thereby improving interpretability and the reliability of subsequent biological inferences. This feature is particularly valuable given the inherent variability and noise in EM segmentation datasets, where ambiguous or borderline cases are common.
The implications of NEURD’s capabilities extend beyond cell-type classification. Automated proofreading and feature extraction reduce the massive time and labor bottlenecks traditionally associated with connectomics studies. By integrating machine learning techniques with expert-verified segmentation datasets, the pipeline paves the way for scalable, high-throughput connectomic analyses across different brain regions, species, and developmental stages.
Furthermore, the modular design of NEURD, combining morphological segmentation with graph-based learning, opens avenues for exploring the structural basis of neural computations and circuit motifs. By accurately categorizing neurons according to their dendritic and synaptic profiles, neuroscientists can better correlate cell types with functional roles, physiological properties, and disease relevance in health and pathology.
Notably, this work underscores the complementarities between high-resolution structural data and computational modeling, illustrating how data-driven frameworks can bridge gaps in biological understanding. As volumetric datasets grow in size and complexity, frameworks like NEURD will be instrumental in facilitating discoveries that were previously unattainable due to the formidable data processing demands and annotation challenges.
In summary, NEURD represents a significant methodological advance in connectomics, providing an automated, interpretable, and highly accurate approach for cell-type classification based on electron microscopy data. Its ability to decode cellular identity from dendritic structures alone, and its demonstrated generalizability across datasets, mark a transformative step toward comprehensive brain mapping and understanding the organizational principles that underlie neural circuits.
Subject of Research: Automated proofreading and feature extraction in connectomics for accurate cell-type classification using electron microscopy data.
Article Title: NEURD offers automated proofreading and feature extraction for connectomics.
Article References:
Celii, B., Papadopoulos, S., Ding, Z. et al. NEURD offers automated proofreading and feature extraction for connectomics. Nature 640, 487–496 (2025). https://doi.org/10.1038/s41586-025-08660-5
Image Credits: AI Generated
DOI: https://doi.org/10.1038/s41586-025-08660-5
Tags: automated analysis of neural datasetsautomated proofreading toolscellular and subcellular brain functionscomputational approaches in connectomicselectron microscopy in neuroscienceEM data for neural stratificationfeature extraction in connectomicslarge consortia contributions to neurosciencemicroscopy-derived neural featuresmorphology and transcriptomics integrationneural circuit mappingneuron classification methods