Abstract—Objective: Large-scale and automated detection of fluorescent microscopic synaptic images are essential for the understanding of brain function and disorders at the molecular level. However, the quantification of synapses from fluorescent images is challenging due to low signal-to-noise (SNR) and non-synaptic background artefacts. This calls for new tools to be developed for an automatic, high-throughput and robust synapse image segmentation. Methods: we proposed an automatic synapse segmentation framework using a deep learning method based on a modified U-Net++ and Gabor-based anisotropic diffusion (GAD). The modified U-Net++ was used to segment the non-synaptic regions, while the multiplicative Poisson noise was suppressed and the edge of the synapses was enhanced by the GAD filter. Thereafter, the synapses were segmented by a thresholding method. Results: The non-synaptic regions were segmented precisely, and the Dice coefficient and Jaccard similarity were 0.833 and 0.719. Our model for synapse segmentation reduced the interference from the non-synaptic tissues and Poisson noise and yielded automatic and accurate segmentation of synapses. Conclusion: We have proposed an automatic segmentation framework that can accurately segment non-synaptic and synaptic tissues, which may have the potential to automate the quantitative analysis of synapses.

Keywords: synapse, image segmentation, Gabor-based anisotropic diffusion, U-Net++