Graph partitioning algorithms with biological connectivity decisions for neuron reconstruction in electron microscope volumes

Abstract

Neuron reconstruction algorithms used in electron microscope volumes have received increasing attention in recent years. Most current methods are highly reliant on neuron membrane boundary evidence without considering biological plausibility. In this investigation, we present a novel neuron reconstruction framework via the fusion of a global optimization goal and biologically inspired priors. We encode the 3D instances of synapses and mitochondria as two types of constraints to allow for the direct inclusion of non-local connectivity information in the neuron segmentation. Moreover, a flexible decision procedure is designed to retain high-confidence priors to deal with the possible influence of the upstream ultrastructure error. We construct the constrained graph partitioning model and adapt two greedy algorithms with the polynomial time complexity to solve the proposed model. We perform comparative studies on several public datasets and demonstrate that the decision of ultrastructural connectivity constraints contributes to significant improvements over existing hierarchical agglomeration algorithms. The ablation studies of ultrastructures from different recognition accuracy suggest the generality and applicability of the proposed method.