A morphology-spatial awareness network for neuronal classification and retrieval

The study of neuronal morphology is vital for brain research. Traditional morphometric methods rely on global or handcrafted features, while current deep learning struggles to capture key multi-scale features and spatial structures, limiting representation of complex morphology. To address these challenges, we propose an end-to-end morphology-spatial awareness network (MSA-Net) based on point clouds. By representing 3D neuronal morphologies as point clouds, MSA-Net faithfully preserves their intrinsic geometric information. Specifically, it integrates a spatial-aware clustering (SAC) attention mechanism to adaptively capture discriminative features from critical morphological regions, and further proposes a global position-local feature interaction (GPLFI) module to facilitate the synergy between global positional embeddings and local point features, ultimately generating compact and robust low-dimensional representations for individual neurons. Extensive evaluations on two datasets demonstrate that MSA-Net exhibits strong discriminative capability in characterizing complex neuronal morphologies and consistently outperforms state-of-the-art methods in both classification and retrieval tasks, thereby establishing a robust paradigm for automated, large-scale neuronal morphology analysis. Our code is available at https://github.com/Jerry01902/MSA-Net .