An anchor-free instance segmentation method for cells based on mask contour

Abstract

Detection and segmentation of cells can be of great significance to the further quantitative analysis of biomedical research in the field of biomedical engineering. Especially, it is a serious challenging task for some microscope imaging devices with limited resources owing to a large number of learning parameters and computational burden when using the detect-then-segment two-stage strategy. In this work, an anchor-free instance segmentation method is proposed for cells based on mask contour. Specifically, an anchor-free network framework is firstly designed for instance segmentation by replacing the standard convolution in the contour generation branch with a deformable convolution. Then, these key contour points are linked to generate coarse contours of cells by using a Graham algorithm. Thirdly, the obtained coarse contours are used for regressing and approaching the ground truths in polar coordinates under the supervision of the Mask loss function. Finally, a series of comparison experiments are conducted to verify the effectiveness of the proposed methods on various datasets. The results show that the proposed method can obtain a better trade-off between recognition performance and computing efficiency, and it can surpass the existing one-stage SOTA methods in the Dice coefficient while maintaining higher computational efficiency on different datasets. Even with a two-stage instance segmentation method like Mask R-CNN, the proposed method can only obtain slightly lower Dice coefficients but with much higher FPS.