Neuron Segmentation Based on CNN with Semi-Supervised Regularization

Abstract

Neuron segmentation in two-photon microscopy images is a critical step to investigate neural network activities in vivo. However, it still remains as a challenging problem due to the image qualities, which largely results from the non-linear imaging mechanism and 3D imaging diffusion. To address these issues, we proposed a novel framework by incorporating the convolutional neural network (CNN) with a semi-supervised regularization term, which reduces the human efforts in labeling without sacrificing the performance. Specifically, we generate a putative label for each unlabeled sample regularized with a graph-smooth term, which are used as if they were true labels. A CNN model is therefore trained in a supervised fashion with labeled and unlabeled data simultaneously, which is used to detect neuron regions in 2D images. Afterwards, neuron segmentation in a 3D volume is conducted by associating the corresponding neuron regions in each image. Experiments on real-world datasets demonstrate that our approach outperforms neuron segmentation based on the graph-based semisupervised learning, the supervised CNN and variants of the semi-supervised CNN.