Enhancing Neural Segmentation on Low-Quality SEM Images Using Self-Attention Based Fusion Module with Auxiliary Deblurring Network

The connectomics and morphology of nerves are essential to basic neuroscience since the neuron system is crucial in the human body. To reconstruct the morphological structure of nerve cells, highly robust segmentation methods are necessary for biological neural connectome analysis. However, current continuous imaging techniques, scanning electron microscope (SEM), inevitably generate low-quality images from automated data acquisition, making neural segmentation a challenging task. Deep learning has made progress in large-scale instance segmentation, but current methods have several limitations when processing blurry images. In this paper, we propose an adaptive attention-based fusion with an auxiliary deblurring decoding network and a main segmentation decoding network to enhance neural segmentation on low-quality SEM images. Our method uses dual task learning to split the complex task into two parts: segmentation and deblurring. The model is designed in an end-to-end structure with a shared encoder and two separate decoders, where it can output both segmentation and deblurring results simultaneously. The proposed method outperforms other methods on two dataset and gains the lowest Voi and A-Rand. The proposed method improves the robustness and accuracy of neural segmentation by enhancing low-quality SEM images. Furthermore, this architecture provides a potential solution against low-quality effects.