Segmenting Neuronal Structure in 3D Optical Microscope Images via Knowledge Distillation with Teacher-Student Network

Three-dimensional (3D) volumetric neural image segmentation is crucial to reconstructing accurate neuron structures. However, due to the structural complexity of neurons and the diverse imaging qualities of the microscopes, it is challenging to achieve both accuracy and efficiency. In this paper, we propose a teacher-student learning framework for fast neuron segmentation. The segmentation inference is performed using a light-weighted student network which benefits from knowledge distillation of a teacher network with a higher capacity. Evaluated on the Janelia dataset from the BigNeuron project, our proposed framework achieves competitive performance for segmentation accuracy while reducing the computational cost to facilitate large-scale processing.

1. Introduction

3D neuron reconstruction, also known as 3D neuron tracing, is a procedure of rebuilding the 3D morphology of neurons using optical microscope images [1, 2]. Reconstruction of neuronal structures such as curvilinear arbours and blobshaped soma is crucial to determining a neuron’s connectivity and, furthermore, its functionality in the neural circuit. Existing 3D neuron reconstruction algorithms can be summarized as a pipeline consisting of pre-processing the optical image, tracing the neuron structure, and post-processing the traced image [3]. Image segmentation, as one main part of the pre-processing phase [4], generates a segmented map where each voxel of an optical image is classified as a foreground or background cluster. Tracing then becomes easier and helps improve both the efficiency and accuracy of the whole procedure. Fig. 1.

Noise patterns and discontinuous structures of 3D neuron images.