Deep residual contextual and subpixel convolution network for automated neuronal structure segmentation in micro-connectomics

Abstract

Background and objective: The goal of micro-connectomics research is to reconstruct the connectome and elucidate the mechanisms and functions of the nervous system via electron microscopy (EM). Due to the enormous variety of neuronal structures, neuron segmentation is among most difficult tasks in connectome reconstruction, and neuroanatomists desperately need a reliable neuronal structure segmentation method to reduce the burden of manual labeling and validation.

Methods: In this article, we proposed an effective deep learning method based on a deep residual contextual and subpixel convolution network to obtain the neuronal structure segmentation in anisotropic EM image stacks. Furthermore, lifted multicut is used for post-processing to optimize the prediction and obtain the reconstruction results.

Results: On the ISBI EM segmentation challenge, the proposed method ranks among the top of the leader board and yields a Rand score of 0.98788. On the public data set of mouse piriform cortex, it achieves a Rand score of 0.9562 and 0.9318 in the different testing stacks. The evaluation scores of our method are significantly improved when compared with those of state-of-the-art methods.

Conclusions: The proposed automatic method contributes to the development of micro-connectomics, which improves the accuracy of neuronal structure segmentation and provides neuroanatomists with an effective approach to obtain the segmentation and reconstruction of neurons.

Keywords: Deep learning; Electron microscopy; Micro-Connectomics; Neuronal structure segmentation; Subpixel convolution.