Abstract
Semi-supervised learning offers a cost-effective approach for neuron segmentation in electron microscopy (EM) volumes. This technique leverages unlabeled data to regularize supervised training for robust neuron boundary prediction. However, distribution mismatch between labeled and unlabeled data, caused by limited annotations and diverse neuronal structures, limits model generalization. In this study, we develop a distribution-aware pipeline to address the inherent mismatch issue and enhance semi-supervised neuron segmentation in EM volumes. At the data level, we select representative sub-volumes for annotation using an unsupervised measure of distributional similarity, ensuring broad coverage of neuronal structures. At the model level, we encourage consistent predictions across mixed views of labeled and unlabeled data. This design prompts the network to align feature distributions and learn shared semantics. Experiments on diverse EM datasets demonstrate the effectiveness of our method, which holds the potential to reduce proofreading demands and accelerate large-scale connectomic reconstruction efforts.