Edge-Convolution Latent Inference for Brain Tractography Embedding

Brain connectivity studies increasingly use diffusion MRI tractography to build whole-brain structural networks as tractograms with millions of streamlines. But streamlines are variable-length polylines with irregular point spacing, which makes geometric processing and distance computation challenging, and complicates downstream tasks such as artifact detection, bundle identification, streamline manipulation, and tractogram comparison. To address this issue, common fixes either resample streamlines to a fixed length (distorting geometry) or learn distance-based embeddings (e.g., dissimilarity representations, autoencoders) that require aggressive resampling and often generalize poorly across subjects. Deep models also treat streamlines as unordered point clouds, discarding the edge relationships that define anatomical trajectories. We present ECLIPSE (Edge-Convolution Latent Inference for Path/Stream Embedding), a geometry-aware framework that embeds streamlines by explicitly modeling point-to-point edges along the path. Using edge convolution to propagate information between neighboring vertices, ECLIPSE preserves sequential structure while capturing smoothness, curvature, and local anatomical context. Across datasets spanning different step sizes and tractography algorithms, ECLIPSE generalizes robustly to geometric and tractography variability, yielding more faithful embeddings for downstream tasks – including bundle identification and clustering – than existing streamline embedding methods.