An Automatic Tracking Method For Multiple Cells In Microscopic Sequence Based On Faster R-CNN

In many applications of biomedicine,the detection and automatic tracking of multiple cancer cells under microscopic sequence images are the basis of recording and analyzing the life cycle activity of cancer cells,which are of great significance for drug development,disease diagnosis and treatment.As a new microscopy technique,phase contrast microscope can be used to observe living cells without staining them,which is helpful for monitoring the life cycle of cancer cells and subsequent dynamic analysis.But the phase contrast microscopic images of cancer cells have some disadvantages,such as low contrast,variable shape,partial occlusion,indefinite number and high cell density,and it is very difficult to solve these problems by adopting some traditional tracking methods.In this work,a tracking-by-detection method is proposed for cancer cells.Firstly,the initial detection of cancer cells was conducted based on Faster R-CNN,and the obtained detection results are analyzed so as to optimize the detection algorithm and improve the detection accuracy.Then,by analyzing the detection results,an optimized detection algorithm is proposed to solve the existing problems and improve the detection accuracy.And a multi-target tracking algorithm is proposed by utilizing the above detection results.Finally,basing on the tracking results and according to the specific characteristics of cancer cells,the tracking results are optimized to improve the tracking efficiency.The main work and innovation points of this thesis are listed as follows:1.To deal with the problem of high density of cancer cells,a detection algorithm is proposed by fusing density estimation based on Faster R-CNN and a multi-task loss function is construted to optimize the detection results.2.To solve the problem of poor feature of cancer cell images under phase contrast microscope,a multi-feature fusion algorithm is presented for cells tracking by fusing the target location information,velocity information and conv features from target detection result to multi-feature descriptors,and it can improve the characterization of features.3.To reduce the high computational cost in the tracking stage,a hierarchical cell tracking algorithm is put forward by classifying cells into lazy and active according to their active degree,then each cell is tracked according to its category,such method can improve the tracking efficiency.4.To eliminate the FN and FP in tracking stage,some context information is fused into the tracking algorithm to improve tracking results.The above algorithms are tested on two public datasets,which are obtained from ISBI2015 competition,and on the T24 dataset of bladder cancer,which come from Cambridge University.And then the experimental results show that the performance of the optimized detection algorithm and optimized tracking algorithm are improved in each dataset.