Nested U-Net Architecture Based Image Segmentation for 3D Neuron Reconstruction

Abstract

Digital reconstruction of neurons is a critical step in studying neuronal morphology and exploring the working mechanism of the brain. In recent years, the focus of neuronal morphology reconstruction has gradually shifted from single neurons to multiple neurons in a whole brain. Microscopic images of a whole brain often have low signal-to-noise-ratio, discontinuous neuron fragments or weak neuron signals. It is very difficult to segment neuronal signals from the background of these images, which is the first step of most automatic reconstruction algorithms. In this study, we propose a Nested U-Net based Ultra-Tracer model (NUNU-Tracer) for better multiple neurons image segmentation and morphology reconstruction. The NUNU-Tracer utilizes nested U-Net (UNet++) deep network to segment 3D neuron images, reconstructs neuron morphologies under the framework of the Ultra-Tracer and prunes branches of noncurrent tracing neurons. The 3D UNet++ takes a 3D microscopic image as its input, and uses scale-space distance transform and linear fusion strategy to generate the segmentation maps for voxels in the image. It is capable of removing noise, repairing broken neurite patterns and enhancing neuronal signals. We evaluate the performance of the 3D UNet++ for image segmentation and NUNU-Tracer for neuron morphology reconstruction on image blocks and neurons, respectively. Experimental results show that they significantly improve the accuracy and length of neuron reconstructions.