Voxel-Wise Cross-Volume Representation Learning for 3D Neuron Reconstruction

Automatic 3D neuron reconstruction is critical for analysing the morphology and functionality of neurons in brain circuit activities. However, performance existing tracing algorithms hinged by low image quality. Recently, a series deep learning based segmentation methods have been proposed to improve quality raw optical stacks removing noises restoring neuronal structures from low-contrast background. Due variety lack large datasets, most current models rely on introducing complex specially-designed submodules base architecture with aim encoding better feature representations. Though successful, extra burden would be put computation during inference. Therefore, rather than modifying network, we shift our focus dataset itself. The encoder-decoder backbone used attends only intra-volume voxel points learn structural features but neglect shared intrinsic semantic voxels belonging same category among different volumes, which also important expressive representation learning. Hence, utilise scarce dataset, propose explicitly exploit such through novel voxel-level cross-volume paradigm basis an model. Our method introduces no cost Evaluated 42 images BigNeuron project, demonstrated ability original model further enhancing performance.