Research On Multi-cell Tracking For Jointing Detection And Segmentation

Cell tracking is one of the important research topics in the field of microscope imaging.Cells have morphological changes and mitosis.Therefore,cell tracking is more challenging than general object tracking.The current mainstream cell tracking methods are usually based on the detection or segmentation of cells.These methods perform well in cases where the distinction between cells is significant or sparse.However,in the scene of highly dense cells,these methods often have poor tracking results due to missing cells.Based on U-Net and related tracking methods,we improve and develop a novel multi-cell tracking method,which can improve performance of multi-cell tracking in the scene of highly dense cells.For serious problems of missed in cell detection,a method that only identifies the cell centroid region is proposed based on U-Net.It makes cell localization and counting more accurate.The multi-frame input method can make the network to learn spatio-temporal information of sequence images.Without changing the structure of U-Net,the multi-frame input method can improve the detection accuracy of mitotic cells,and further improve the performance of the cell mitosis behavior detection.The attention mechanism is integrated into U-Net to adaptively calibrate the output of network layers.With advantages of the strong global context modeling capabilities by NL block and the lightweight computing by SE block,a global context information response self-calibration block is designed,named GCR Block.GCR Block is mainly composed of global context information extraction and pixel-level response self-calibration.Experimental results show that,after integrating into GCR block,cell centroid detection results generated by U-Net can significantly improve the performance of multi-cell tracking.In terms of improving the tracking framework,a multi-cell tracking method is proposed,which jointly uses cell centroid detection and cell segmentation,namely JCDSMT.Through the target association of cell centroids,cell centroid motion trajectories are established,and the number of trajectory breaks is reduced.In the target association,Kalman filter is used to predict the motion state of each centroid trajectory and matching is performed based on multiple distance cost terms.According to the state and trajectories information of cells,a local cell state matrix is established to detect the cell mitosis behavior,thereby building up the cell lineage.Combined with the termination information of centroid motion trajectories,the cell apoptotic behavior is detected.Combined with results of cell primary segmentation,cell further segmentation is performed.Cell further segmentation provides a unique mask for each cell centroid to obtain final multi-cell tracking results.In cell further segmentation,a method for accelerated calculation using the Voronoi polygon method is proposed.A variety of datasets are used to verify the robustness of JCDSMT.In the official platform evaluation of Cell Tracking Challenge,JCDSMT achieved the state-of-the-art performance on dataset He La.In the evaluation of using CMU dataset,the tracking performance of JCDSMT far exceeds the tracking method LM,which is based on cell segmentation.Experimental results have shown that JCDSMT achieves stable and accurate performance of multi-cell tracking in the scene of highly dense cells.