SimCP-NS: Similarity-based Copy Paste for Semi-Supervised 3D EM Neuron Segmentation.

Abstract

Semi-supervised neuron segmentation in 3D electron microscopy (EM) is important for connectomics because it enables accurate neuron reconstruction while reducing dependence on costly manual annotations. However, existing approaches remain limited by the distribution mismatch between the small labeled dataset and the much larger unlabeled dataset, where the labeled data fails to adequately capture the true data distribution, leading models trained on them to generate low-quality segmentation masks and ultimately degrading overall performance. To address this issue, we propose a similarity-based copy-paste for semi-supervised 3D EM neuron segmentation (SimCP-NS) method, which employs a similarity-based copy-paste strategy to exchange the least similar labeled and unlabeled sub-volumes, thereby enriching data diversity and mitigating distribution mismatch. A teacher network is first pre-trained on unlabeled volumes to capture structural priors, which subsequently guides the student segmentation network. During student training, the similarity-based copy-paste mechanism generates hybrid samples and constructs supervision targets by fusing teacher-generated pseudo-labels with ground-truth affinity maps, optimized via mean squared error loss. Invariant representation learning is further integrated to enhance robustness of the proposed method. Extensive experiments demonstrate the superior performance of SimCP-NS over existing 3D EM neuron segmentation methods. Codes and other supporting materials are provided in the Supplementary Material.