Automated Neuron Tracing with Imitation Learning

Abstract

Reconstructing neuronal morphology from large 3D microscopy volumes is essential for quantitative neuro-science but remains challenging due to noise, low contrast, and complex branching geometry. We present an automated neuron tracing method that formulates reconstruction as a sequential decision-making problem and trains a 3D convolutional policy via imitation learning. The key technical contribution is an online expert-action retrieval scheme that derives locally valid continuation directions directly from a gold-standard SWC tree, including at bifurcations. Combined with DAgger-style dataset aggregation, the policy is trained under its own state distribution without requiring handcrafted reward functions or a separate segmentation-to-skeleton postprocessing stage. We package the approach into an end-to-end tracing pipeline with optional path refinement operations and tools for quantitative evaluation and interactive editing. On the BigNeuron Gold166 benchmark, our method achieves state-of-the-art reconstruction accuracy, outperforming all competing challenge methods under both morphology-based aggregate metrics and geometric distance comparisons.