Cell instance segmentation is a critical task to perform for the quantitative analysis of 3D live-cell images. Existing studies mostly apply a region proposal-based approach to instance segmentation of microscopy images. However, they often fail to detect cells in 3D live-cell images, which have complicated and heterogeneous shapes, often closely linked to the neighborhood cells. A different approach based on point proposal methods is more robust in handling complex shapes than the box proposal. These methods take an image and a proposed point in the form of its location (x; y) as input and generate a mask for an object that includes the point. They also show that the model can improve the prediction by utilizing negative point proposals chosen from false-positive areas. In this paper, we propose a novel cell instance segmentation approach based on point proposal for 3D cell imaging. Different from existing work, however, our model utilizes the nuclei of cells as point proposal and employ them as positive and negative point proposals. We constructed the 3D NIH3T3 dataset for training and evaluation, and examine the proposed model qualitatively on three independently gathered cells; HeLa, A549, and MDA-MB-231. Our model exhibits superior quantitative results; moreover, compared to previous methods, it properly predicts cell lines, which are not even well-annotated during training.
|