NeMoCo: Self-supervised contrastive learning for ultrastructural 3D neuroscience morphologies

Volume electron microscopy (EM) now enables nanometric-scale 3D reconstructions of neural tissue, opening the door to quantitative, morphology-driven neuroscience beyond connectivity alone. While previous studies relied on handcrafted descriptors and classical machine learning for morphology analysis, recent progress in deep learning for 3D shape understanding offers new opportunities to learn robust, task-specific representations directly from geometric data. In this paper we present NeMoCo, a geometry learning framework that targets the key practical bottleneck in connectomics and ultrastructural analysis: the scarcity and cost of dense expert annotations for the long tail of neurite and organelle phenotypes. NeMoCo formulates representation learning for EM-derived neurite meshes in a self-supervised Momentum Contrast (MoCo) style. We use DiffusionNet (Sharp et al., 2022) as a mesh encoder with intrinsic spectral descriptors (HKS) and train with a momentum-updated teacher encoder and a large memory bank of negatives. To learn invariances that are essential in practice, we generate paired geometric views via controlled affine transformations and resolution changes (including mesh decimation), encouraging embeddings to be stable under nuisance variability while remaining discriminative. We provide an extensive study of augmentation strength and temperature, and evaluate learned representations through frozen retrieval and non-parametric classification (frozen kNN), as well as downstream supervised fine-tuning under limited labels. NeMoCo demonstrates that MoCo-style self-supervision yields robust neurite morphology embeddings on EM meshes, improving label-efficiency and offering a scalable foundation for retrieval, clustering, and phenotype discovery in ultrastructural neuroscience. All the data and the code used for NeMoCo are available at https://github.com/Uzshah/NeMoCo.