Research On Cell Tracking Based On Deep Seed And Local Graph Matching

Shoot apical meristem is the most important tissues,because it supplies nutrients to cells in other parts of the plant and differentiates into plant organs.Research on the apical meristem cells helps researchers understand the cell’s tissue structure,growth rules and physiological functions,which will be benefit to medical research and drug development.Plant cells are physically adjacent to each other through a shared cell wall,so the relative position between cells changes little.The early cell local matching algorithm takes advantage of the fact,and used an iterative search strategy to continuously expand the matching cells from the seed cell through the neighborhood structure.However,the shape of the cells changes greatly over time,so this strategy is easy to accumulate errors.Once the seed point is not selected correctly,or an error occurs in the matching expansion,it will lead to a series of errors in subsequent matching.In order to enhance the robustness of seed cell detection and thus increase the accuracy of cell tracking,a cell tracking algorithm based on Deep Seed and local graph matching was proposed.The algorithm process is: 1.Segmenting the cell image by using the watershed algorithm.2.Building a cell similarity learning network based on convolutional neural network,and the trained cell similarity learning network model is used to extract the cells’ deep feature.3.Extracting the cells’ local triangle graph feature in the cell segmentation image.4.Combining the deep feature and the local triangle graph feature to establish the similarity matrix.Deep Seed is the cell pair correspondence of the maximum value of the similarity matrix.5.Starting from the Deep Seed,all cell pairs are matched by neighborhood correspondence growing process.The core of the paper is to establish a similarity model by combining local graph matching and cell similarity learning network to robustly find the Deep Seed and perform subsequent tracking processes.Among them,the cells’ local triangle feature model uses the stability of the cells’ spatial-temporal contextual information(cell area ratio,cell edge length ratio,cell angle).The cells’ feature learned by the cell similarity learning network proposed is used to establish cells’ deep feature model.Compared with traditional cell tracking methods,the experimental results show that this algorithm is applied to the unregistered image sequences and large interval image sequences of shoot apical meristem cells,and the tracking accuracy is high.