Deep Learning Approaches for Large-Scale 3D Neuron Segmentation
by
OHSA/B17
Abstract:
Automated neuron segmentation from large-scale electron microscopy (EM) volumes is a central challenge in brain mapping. In the ALBUM project, we address artifacts such as section loss, folds, and limited z-resolution that hinder segmentation accuracy, with the goal of reconstructing higher-quality 3D volumes. Building on this context, I will present Cross-dimension Affinity Distillation (CAD), a recent method for efficient 3D EM neuron segmentation. CAD leverages lightweight 2D CNNs while distilling inter-section dependencies from a 3D teacher network and introducing feature grafting to enhance knowledge transfer. This strategy achieves state-of-the-art segmentation accuracy at a fraction of the computational cost, offering a promising direction for scalable analysis of EM data.
The Laboratory for Simulation and Modeling
SDSC hub at PSI