3D Neuron Segmentation Based on 3D DSAC U-Net

摘要

In the brain, the nerve fibers of adjacent neurons are intertwined. If the three-dimensional neuron image is directly tracked and reconstructed, it is easy to be interfered by nerve fibers and other noises. Therefore, the basic work of reconstructing or tracing 3D neuron structure is to segment the target neuron from neuron images containing nerve fiber entanglement. Today, deep learning has a wide range of applications in the field of medical image segmentation, and the most representative segmentation model is U-Net. However, the traditional U-Net model has some shortcomings, for example, the jump connection directly transmits the features to the corresponding up-sampling layer without processing, which may cause edge misjudgment. In this article, we propose an improved method based on 3D U-Net. We redesign the jump connection, add the residual path, enhance the low-resolution information, extract the context information through the hole convolution and splice it to the corresponding up-sampling layer, and add an upper jump connection to supplement the loss of deep learning Detailed information. Compared the proposed model with other networks, we come to conclusion that our proposed model significantly improved the accuracy of neuron segmentation.