VENturing into machine learning for the morphological analysis of von Economo neurons

Von Economo neurons (VENs) are a specialized type of large, highly elongated projection neurons located in specific cortical regions. Despite their implication in higher-order cognitive functions and psychiatric disorders in humans, consistent and objective identification criteria for VENs remain lacking. We analyzed 761 digitally reconstructed neurons from the NeuroMorpho.Org database. We applied six supervised machine learning algorithms and a convolutional neural network with Grad-CAM visualization to classify the reconstructions into VENs and pyramidal neurons. Variable importance was evaluated using information-driven and expert-based selection. We compared the classifications made by machine learning algorithms to the reconstructions' original labels. Reconstructions misclassified by the classifier models were further examined by a neuroanatomy expert. Machine learning models generally achieved high classification accuracy. Morphometric features such as dendritic length and number of stems emerged as some of the key discriminators. Expert ratings only partially aligned with machine findings, and there was low agreement between experts. Most misclassifications made by the classifier models were attributable to reconstruction artifacts or ambiguous morphology rather than model limitations. Our findings demonstrate the utility of combining machine learning with expert insight for distinguishing VENs from pyramidal neurons. While soma shape remains important for the characterization of VENs, classifier models revealed that dendritic architecture may be equally as specific and could help distinguish between borderline cases. This framework offers a replicable, data-driven method for studying VENs and can be utilized for future research on their distribution and function.