DPA: Deep Parcellation Analysis for Human Trait Prediction

Abstract Mapping individual differences in white matter organization is a central goal of human brain mapping. Atlas-based structural connectome analyses are widely used but are sensitive to anatomical assumptions and may miss fine-grained connectivity patterns. Data-driven fiber clustering mitigates some limitations but typically relies on streamline endpoints, overlooking rich geometric information along full trajectories. In this paper, we introduce Deep Parcellation Analysis (DPA), a geometry-aware framework for white matter parcellation and trait association. DPA samples multiscale geometric features along entire fiber streamlines and encodes them with a Multiscale Fiber AutoEncoder (MFAE) to produce compact latent representations. These enable clustering that preserves intrinsic fiber geometry independently of anatomical atlases, supporting interpretable mapping of connectivity patterns linked to individual differences. Applied to diffusion magnetic resonance imaging data from the Human Connectome Project Young Adult (HCP-YA) cohort, DPA captures white matter substrates underlying language-related cognitive traits. It achieves comparable or superior predictive performance relative to atlas-based and principal parcellation approaches while revealing biologically plausible fiber pathways. Critically, DPA identifies trait-relevant connections missed by endpoint-based methods, emphasizing the importance of full streamline geometry. Overall, DPA demonstrates that deep, geometry-informed parcellation provides a powerful, interpretable approach for linking white matter organization to behavior, advancing structure–function studies in human brain mapping.