Research On Segmentation And Tracking Method Of Micro Cellular Video

Most of the disease is related to the cell behaviors in medicine. Cell segmentation and cell tracking are important means to research cell behaviors. The traditional approach is based on artificial ways, it not only need a lot of manpower, but also need add dye in cells before tracking, the dye is a chemical substance, affects the cell physiological activity, and the affects the results. Therefore, it is of great significance to study the cell behavior by using the theory of computer vision to complete cell segmentation and cell tracking.Aiming at cell segmentation and cell tracking, the following work is finished in this paper:1)This paper proposes a four-stage method to segment the touching cells in image, which based on the multi-scale Laplacian-of-Gaussian filter to detect cell centers. Our method involves, the threshold method to extract the cellular region in image; a multi-scale Laplacian-of-Gaussian filter to detect cell centers; a method based on the City-Block distance to locate the junction part in touching cells, then by removing the low-gray pixels in junction part or the nearest neighbor method to get a coarse segmentation results; last, by adding the gradient constraint term improve the active contour model to optimize the result, which makes the final segmentation result close to the real cellular boundary. The proposed method is quantitatively validated over MCF-10 A non-transformed breast cell images. The results show that our approach provides more accurate segmentation especially when cell-cell-contacts.2) In order to deal with the complicated situation of cell division, disappearance and rebirth, the cell tracking problem is transformed into a classification problem, then based on supervised learning and active learning finish the cell tracking. In this method, by mining of cell characteristics and using GBDT model training, can obtain a good tracking result in unknown samples. Aiming at the problem that microscopic cell video hasn't a lot of labels, we using the active learning method to solve it. By developing some strategy to select samples and label them, we can get a high rate of accuracy by labeling very few samples.3) PHD(Probability Hypothesis Density) filter is based on the random finite set statistical filtering method, which can track the change number targets, so it is suitable for cell tracking. Under the probability hypothesis density filter tracking framework, by analyzing the characteristics of the cells in the microscopic video images, the cell center location, gray mean level and local scene image entropy features are used to characterize the kinematic and image characteristics of the cell, and then automated cell tracking. The experimental results show that this method can track a lot of cells effectively, and has a higher tracking accuracy compared with only using the single dynamic feature.