Selective Labeling Meets Semi-Supervised Neuron Segmentation

ABSTRACT Semi-supervised learning offers a cost-effective approach for neuron segmentation in electron microscopy (EM) volumes. This technique leverages extensive unlabeled data to regularize supervised training for more robust predictions of neuron affinities. However, the distribution mismatch between labeled and unlabeled datasets, arising from limited annotations and the diversity of neuronal patterns, impedes the generalization of semi-supervised models. In this paper, we develop a dual-level pipeline to address the inherent mismatch issue and enhance semi-supervised neuron segmentation. At the data level, we propose an unsupervised heuristic to select valuable sub-volumes as labeled data based on distribution similarity in a pretrained feature space, ensuring a representative coverage of neuronal structures. At the model level, we introduce an axial-through mixing strategy into anisotropic neuron segmentation and integrate it into a semi-supervised framework. Building on this, we establish cross-view consistency constraints through intra- and inter-mixing of labeled and unlabeled datasets, which facilitates the learning of shared semantics across distributions while avoiding ambiguity in neuron affinities. Extensive comparative experiments and ablation studies on publicly available datasets demonstrate the effectiveness of the proposed method across different EM modalities and spatial resolutions.