Cell Tracking Across Noisy Image Sequences Via Faster R-CNN and Dynamic Local Graph Matching

Abstract

Automated tracking of cells in time-lapse live-imaging datasets of developing multicellular tissues is gaining popularity in developmental biology for understanding the cell growth dynamics. The tracking of plant cells across noisy microscopy image sequences is very challenging, because plant cells in noisy region cannot be correctly segmented and cause serious errors in subsequent cell tracking procedure. In this paper, we present to track plant cells across noisy images using a tracking method which is based on Faster R-CNN and dynamic local graph matching. Faster R-CNN is employed to detect cells in noisy images, and it is improved by cell characteristic prior bounding box design and soft non-maximum suppression strategies. Then a dynamic local graph matching model is proposed to track the detected plant cells, by exploiting the cells’ tight spatial and temporal contextual information. It tends to prevent the cell matching error accumulation by selecting the most similar cell pair in the dynamically growing neighbor set of matched cells. Compared with the existing tracking methods for plant cells, the experimental results show that the proposed method can greatly improve the tracking accuracy.