NeuroShape: exploiting neural architectures for shape analysis of ultrastructural 3D neuroscience morphologies

Recent advances in volume electron microscopy (EM) enable nanometric-scale 3D reconstructions of neural tissue, providing unprecedented opportunities for studying cellular and subcellular morphology in neuroscience. The geometry of structures such as nuclei, neurites, and organelles can encode phenotypic information relevant to both functional specialization and pathological conditions, and thus represents a valuable complement to connectivity-based approaches in connectomics. 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 work we present NeuroShape, a first exploration of modern deep learning methods for shape analysis of ultrastructural 3D neuroscience morphologies. We introduce two annotated datasets derived from EM reconstructions: one of nuclei envelopes, and one of neurites and neural organelles. We benchmark two state-of-the-art neural architectures for 3D geometry (DiffusionNet [SACO22] and Laplacian2Mesh [DWL*24]) and compare them against traditional feature-based descriptors previously used in neural morphology analysis. Our preliminary results highlight both the feasibility and the challenges of applying deep learning shape analysis techniques in this domain, and we release the datasets as a reference resource for future studies.