Multiscale Kernels for Enhanced U-Shaped Network to Improve 3D Neuron Tracing

Abstract

Digital neuron morphology reconstruction from three-dimensional (3D) volumetric optical microscope images is an important procedure to rebuild the connections and structures of neural circuits. Even though many approaches have been proposed to achieve precise tracing, it is still a challenging task especially when images are polluted by noise or have discontinuity in their neuron structures. In this paper, we propose a new framework to overcome these issues by performing neuron segmentation prior to tracing. Our proposed framework adopts a novel 3D U-shaped convolutional neural network (CNN) with multiscale kernel fusion and spatial fusion to perform the image segmentation. We then perform the iterative back-tracking tracing algorithm on the output of the network. Evaluated on the Janelia dataset from the BigNeuron project, our proposed framework achieves competitive tracing performance.