Cell Instance Segmentation Via Multi-Scale Non-Local Correlation.

For cell instance segmentation on Electron Microscopy (EM) images, state-of-the-art methods either conduct pixel-wise classification or follow a detection and segmentation manner. However, both approaches suffer from the enormous cell instances of EM images where cells are tightly close to each other and show inconsistent morphological properties and/or homogeneous appearances. This fact can easily lead to over-segmentation and under-segmentation problems for model prediction, i.e ., falsely splitting and merging adjacent instances. In this paper, we propose a novel approach incorporating non-local correlation in the embedding space to make pixel features distinct or similar to their neighbors and thus address the over- and under-segmentation problems. We perform experiments on five different EM datasets where our proposed method yields better results than several strong baselines. More importantly, by using non-local correlation, we observe fewer false separations within one cell and fewer false fusions between cells.