Triple-Crossing 2.5D Convolutional Neural Network for Detecting Neuronal Arbours in 3D Microscopic Images

Abstract

The automatic analysis of the 3D optical microscopic images containing neuron cells remains one of the central challenges in the modern computational neuroscience. The varying image qualities make the accurate detection of the curvilinear neuronal arbours elusive. The high computational cost raised by large 3D image volumes also makes the conventional filter-bank learning methods impractical. We present a novel Triple-Crossing (TC) 2.5D convolutional neural network to detect the neuronal arbours in large 3D microscopic volumes with a reasonable computational cost. The network is trained to output a regression map that indicates the presence of the neuronal arbours. The proposed methods can be used as a pre-processing step in an automated neuronal circuit reconstruction pipeline, which enables the collection of large-scale neuron morphological datasets. In our experiments, we show that the proposed methods could effectively eliminate dense background noises and fix the gaps along neuronal arbours. The proposed methods could also outperform the original 2.5D neural network regarding the training efficiency as well as the generalisation performance.