He Fuyun, Huang Xiaoming, Wang Xun, Qiu Senhui, Jiang F, Ling Sai Ho
College of Electronic Engineering, Guangxi Normal University, Guilin, China; Guangxi Key Laboratory of Automatic Detection Technology and Instrument, Guilin, China; Guangxi Key Laboratory of Wireless Wideband Communication and Signal Processing, Guilin, China.
College of Electronic Engineering, Guangxi Normal University, Guilin, China; Guangxi Key Laboratory of Wireless Wideband Communication and Signal Processing, Guilin, China.
Comput Med Imaging Graph. 2021 Apr;89:101871. doi: 10.1016/j.compmedimag.2021.101871. Epub 2021 Feb 23.
Neuron image segmentation has wide applications and important potential values for neuroscience research. Due to the complexity of the submicroscopic structure of neurons cells and the defects of the image quality such as anisotropy, boundary loss and blurriness in electron microscopy-based (EM) imaging, and one faces a challenge in efficient automated segmenting large-scale neuron image 3D datasets, which is an essential prerequisite front-end process for the reconstruction of neuron circuits. Here, a neuron image segmentation method by combining Chan-Vest (CV) model with Deep Boltzmann Machine (DBM) is proposed, and a generative model is used to model and generate the target shape, it take this as a prior information to add global target shape feature constraint to the energy function of CV model, and the shape priori information is fused to assist neuron image segmentation. We applied our method to two 3D-EM datasets from different types of nerve tissue and achieved the best performance consistently across two classical evaluation metrics of neuron segmentation accuracy, namely Variation of Information (VoI) and Adaptive Rand Index (ARI). Experimental results show that the fusion algorithm has high segmentation accuracy, strong robustness, and can characterize the sub-microstructure information of neuron images well.
神经元图像分割在神经科学研究中具有广泛的应用和重要的潜在价值。由于神经元细胞亚微观结构的复杂性以及基于电子显微镜（EM）成像中图像质量的缺陷，如各向异性、边界损失和模糊性，人们在高效自动分割大规模神经元图像3D数据集方面面临挑战，而这是神经元回路重建必不可少的前端过程。在此，提出了一种将Chan-Vest（CV）模型与深度玻尔兹曼机（DBM）相结合的神经元图像分割方法，使用生成模型对目标形状进行建模和生成，并将其作为先验信息添加到CV模型的能量函数中，以添加全局目标形状特征约束，融合形状先验信息辅助神经元图像分割。我们将该方法应用于来自不同类型神经组织的两个3D-EM数据集，并在神经元分割准确性的两个经典评估指标，即信息变异（VoI）和自适应兰德指数（ARI）上始终取得了最佳性能。实验结果表明，该融合算法具有较高的分割精度、较强的鲁棒性，并且能够很好地表征神经元图像的亚微观结构信息。
