A cell tracking technology method based on deep learning and ant colony algorithm

Abstract

The application belongs to the technical field of cell tracking, and in particular discloses a cell tracking method based on deep learning and an ant colony algorithm, wherein multiple cells are accurately detected in a deep learning manner, double determination of the ant colony algorithm and Kalman filtering avoids target loss caused by irregular movement of cells, and accurate determination is made on cells with complex behaviors. The ant colony algorithm cost formula adopted by the application comprehensively considers the centroid information and the contour information, so that not only the centroid coordinate position is considered when interframe matching is performed, but also the cell morphological information is considered, and more accurate tracking effect is achieved when the cell has complex behaviors; the target detection method based on yolov7 has a good detection effect on similar targets of different sizes by using a mosaic enhanced data enhancement method, and the performance is not reduced due to adjustment of the microscope magnification; the method combining yolov7 and U-Net enhances the visibility of the system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cell tracking, in particular to a cell tracking technology method based on deep learning and ant colony algorithm. BACKGROUND

[0002] Cells are the basic building blocks of the human body, and all life activities cannot be separated from cell proliferation, differentiation and migration. Researchers also hope to accurately trace the lineage relationship of diseased cells through cell tracking, so as to determine the cause of the disease. Cell tracking technology is a product of the modern medical field. Traditional cell tracking technology requires operators to observe the microscope with the naked eye, which is a long, high-cost way that greatly wastes labor while easily leading to inaccurate cell tracking results and negative effects of simple and boring work. In order to overcome the shortcomings of manual operation, a fully automatic cell tracking algorithm is developed to replace traditional manual cell tracking operation, which has great research significance.

[0003] The goal of cell tracking is to track all the movement behaviors of cells, including cell movement, proliferation, and migration in and out events. Due to the difficulty of observing and predicting cells, there are many difficulties in realizing cell tracking: 1. Due to the long time required for cell image sequence acquisition, uneven lighting, equipment jitter and other conditions during the acquisition process directly lead to the difficulty of obtaining suitable image sequence dataset; 2. The size of a single cell is too small compared to the whole microscopic image, the signal-to-noise ratio of the image is too low, and the morphological differences of different cells are not obvious, which makes it difficult to detect the cell target in the image sequence; 3. Cells may exhibit complex behaviors such as cell division, sudden acceleration and direction change, cell overlap, and cell migration in and out of the field of view, which greatly affects the reliability of target tracking.

[0004] Therefore, we propose a cell tracking technology method based on deep learning and ant colony algorithm to solve the above problems. SUMMARY

[0005] (I) Technical problems solved

[0006] In view of the shortcomings of the prior art, the present application provides a cell tracking technology method based on deep learning and ant colony algorithm, which solves the problem of various difficulties in realizing cell tracking due to the difficulty of observing and predicting cells.

[0007] (II) Technical solutions

[0008] In order to realize the above-mentioned purpose, the application provides a cell tracking method based on deep learning and an ant colony algorithm, the method accurately detects multiple cells in a deep learning manner, double determination of the ant colony algorithm and Kalman filtering avoids target loss caused by irregular motion of cells, and accurately determines cells with complex behaviors. The method can effectively compensate for the deficiency of artificial cell tracking, effectively reduce the workload of the operator, make the cell tracking method more scientific and objective, and thus improve the tracking accuracy.

[0009] The technical idea of the application is to combine deep learning, image segmentation, Kalman filtering and the ant colony algorithm to build a complete cell tracking trajectory framework, and finally realize accurate tracking of each cell in the cell image sequence. The specific implementation steps include the following:

[0010] S1, label the cell image sequence under the microscope using labellmg, and the label name is the corresponding cell class name, and the storage format is selected as yolo format;

[0011] S2, the labeled cell image sequence is transmitted to the yolov7 network for training to obtain the optimal target detection weight, and the yolov7 network is modified to make it more suitable for the medical image field and output the center point coordinates of the detection frame;

[0012] S3, label the target detection frame obtained by yolov7 using labelme, and transmit the labeled cell segmentation data set to the U-Net network for training to obtain accurate segmentation images of cells in each target detection frame, and record the cell area, cell eccentricity and contour information;

[0013] S4, establish a cost function of the ant colony algorithm, based on cell contour information, cell movement distance and global movement path information, establish a multi-ant colony cell tracking method (MACT), so as to obtain the cell movement trajectory under the given observation through the ant colony algorithm;

[0014] S5, based on the MACT algorithm, the cells that cannot be successfully tracked by the ant colony tracking algorithm are marked as abnormal cells and enter the abnormal processing module;

[0015] S6, introduce Kalman filtering in the MACT algorithm, predict the cell movement direction through the historical information and add the prediction probability equivalent to the pheromone of the ant colony algorithm, and integrate the speed, acceleration and displacement information in the cell historical data to obtain a reasonable prediction value. The prediction obtained by the Kalman filtering is combined with the position under the given observation obtained by the ant colony algorithm to obtain the final cell tracking trajectory;

[0016] S7, further optimizing the existing tracking trajectory through the trajectory remodeling function of the ant colony algorithm in the MACT;

[0017] S8, repeating steps S4 to S7 from the initial frame to the last frame of image, thus completing the tracking of all cells.

[0018] Further, the cost function of the ant colony algorithm in step S4 is established according to the cell contour information, cell movement distance and global movement path information, including:

[0019] The cost function cost of the ant colony algorithm is established as follows:

[0020] Cost(I, j) = aD(I, j) + bE(I, j) + gA(I, j)

[0021] Where I represents the cell to be tracked detected in the current frame, j represents the cell to be tracked in the next frame, a, b, g represent the weights of distance, eccentricity and area respectively, and the sum of the weights is 1.

[0022]

[0023]

[0024] Where n represents the number of all cells to be tracked in the next frame, d i，j represents the distance between the cell to be tracked i and the cell to be tracked j, Eccent i represents the eccentricity of the cell to be tracked i, Area i represents the area of the cell to be tracked i.

[0025] The reciprocal of the loss function is selected as the judgment basis for tracking, which is used to increase the screening effect of the loss function.

[0026] Further, the cells that the ant colony algorithm fails to successfully track in step S5 are marked as abnormal cells, including:

[0027] The cell belonging to the ant colony in the current frame fails to be successfully tracked in the next frame, and the position information of the cell belonging to the ant colony is recorded, which is marked as a cell moving out or an apoptotic abnormal cell.

[0028] The cell belonging to the ant colony in the next frame is not successfully tracked in the current frame, and the position information of the cell belonging to the ant colony is recorded, which is marked as a cell moving in or a division abnormal cell.

[0029] Further, the prediction obtained by the Kalman filter is combined with the position under the given observation obtained by the ant colony algorithm in step S6, including:

[0030] Using the Kalman filtering algorithm, based on the position, speed, acceleration information of the past time, the position prediction value of the current frame is obtained;

[0031] The observation value obtained by the MACT algorithm is set as the observation value;

[0032] The detection result of the MACT algorithm is directly used to obtain the observation value of the current step, the historical data is integrated and the prediction value of the current step is reasonably predicted through Kalman filtering, and finally the tracking result is obtained by the weighted sum of the prediction value and the observation value based on the Kalman filtering gain coefficient after fully considering the reliability of the observation value and the prediction value.

[0033] Further, the trajectory remodeling function of the ant colony algorithm in the step S7 includes:

[0034] Eliminate the trajectory with too small number of ant colonies in the tracking trajectory;

[0035] Design a cell migration and migration module to determine the migrated cell and assign the number and label of the migrated cell.

[0036] (Three) beneficial effects

[0037] Compared with the prior art, the present application provides a cell tracking technology method based on deep learning and ant colony algorithm, which has the following beneficial effects:

[0038] 1、The ant colony algorithm cost formula considering the centroid information and the contour information is adopted, the centroid coordinate position is considered in the inter-frame matching, and the cell morphology information is also considered, so that more accurate tracking effect is realized when the cell behaves complexly.

[0039] 2、The target detection method based on yolov7 is adopted, the data enhancement method of mosaic enhancement makes the tracking system have good detection effect on similar size targets, and the performance will not be reduced due to the adjustment of microscope magnification.

[0040] 3、The method of combining yolov7 and U-Net is adopted, the finally obtained cell tracking trajectory not only contains the trajectory path, but also displays the specific cell segmentation shape, and the visibility of the system is enhanced.

[0041] 4、The ant colony algorithm cost formula considering the centroid information and the contour information is adopted, the centroid coordinate position is considered in the inter-frame matching, and the cell morphology information is also considered, so that more accurate tracking effect is realized when the cell behaves complexly.

[0042] 5、The application records the morphology and position of the removed cells in the cell removal event, and when similar cells appear in the vicinity of the subsequent frame, it is considered that the removed cells have re-migrated, the original number of the cells is assigned, and the rigor and integrity of the cell tracking trajectory are enhanced.

[0043] 6、The application builds a MACT algorithm, combines the ant colony algorithm that can handle very complex combinatorial optimization problems with multi-cell tracking, builds an ant colony for each cell, and realizes the tracking of multiple cells under a microscope through the independent and cooperative working mode of the ant colonies. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 The flowchart of the application is shown in the figure;

[0045] Figure 2 The cell microscopic video image dataset used by the application is shown in the figure;

[0046] Figure 3 The yolov7 detection frame schematic diagram simulated by the application is shown in the figure;

[0047] Figure 4 The detection frame (U-Net network input) obtained by the application is shown in the figure;

[0048] Figure 5 The U-Net segmentation effect image of the application is shown in the figure;

[0049] Figure 6 The tracking result of the microscopic cell image sequence finally realized by the MACT algorithm is shown in the figure;

[0050] Figure 7 The cell lineage tree corresponding to the cell image sequence made by the MACT algorithm is shown in the figure. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0052] EMBODIMENT

[0053] As shown in the figure, one embodiment of the application proposes a cell tracking technology method based on deep learning and ant colony algorithm, which includes the following steps: Figures 1-7

[0054] ​S1, label the cell image sequence under the phase contrast microscope using labelling, and the label name is the corresponding cell class name. The storage format is selected as yolo format.

[0055] S2, the labeled cell image sequence is transmitted to the yolov7 network for training, and the optimal target detection weight is obtained. At the same time, the yolov7 network is modified to make it more suitable for medical image field and output the center point coordinates of the detection frame.

[0056] S21, the labeled cell sequence image dataset is transmitted into the yolov7 network for training, using the mosaic data enhancement method to overcome the target loss caused by different microscope magnifications, selecting the training method with pre-trained weight, and obtaining the optimal target detection result by modifying the parameters and recording the weight;

[0057] S22, open anaconda prompt control terminal, enter the path of yolov7 code in the control terminal, and execute the code as follows: python detect.py–weights.\runs\train\yolov7\weights\best.pt–source.\datasets\findcontact\testImages\--save-txt;

[0058] S23, running (2b) can obtain the yolo format coordinates of each detection frame obtained by detection under the yolov7 weight training;

[0059] S24, add caijian.py code in yolov7, run caijian code to output the corresponding cropped cell image in the given path, and generate the center point coordinates of the detection frame.

[0060] S3, the target detection frame obtained by yolov7 is extracted, labeled using labelme, and the labeled cell segmentation dataset is transmitted to the U-Net network for training, so as to obtain the accurate segmentation image of the cell in each target detection frame, and record the cell area, cell eccentricity and contour information;

[0061] S31, save the target detection frame pictures to be processed in the same folder, and use labelme software to segment and label the cell image;

[0062] S32, use the data conversion script to convert the data format generated by the labeling tool into the data format required by the U-Net model training; the code is as follows: python labelme2seg.py<PATH / TO / LABEL_JSON_FILE>;

[0063] S33, the labeled cell segmentation image data set is transmitted into the U-Net network for training, since the processed image is the target detection part of the original image, since the signal-to-noise ratio is improved a lot, the segmentation effect of the U-Net network is good;

[0064] S34, the related parameters of the U-Net network are adjusted to obtain the optimal segmentation effect, and the optimal parameters are recorded for subsequent actual tracking.

[0065] S4, a cost function of the ant colony algorithm is established, considering cell contour information, cell moving distance and global moving path information, a multi-ant colony cell tracking method (MACT) is established, so that the cell moving trajectory under the given observation is obtained through the ant colony algorithm;

[0066] The cost function of the ant colony algorithm is as follows:

[0067] Cost(I,j)=αD(I,j)+βE(I,j)+γA(I,j)

[0068] Wherein, I represents the detected cell to be tracked in the current frame, j represents the cell to be tracked in the next frame, and α, β and γ represent the weights of distance, eccentricity and area respectively, and the sum of the weights is 1; The values are 0.7, 0.15 and 0.15 respectively;

[0069]

[0070] Wherein n represents the number of all cells to be tracked in the next frame, d i，j represents the distance between the cell to be tracked i and the cell to be tracked j, Eccent i represents the eccentricity of the cell to be tracked i, Area i represents the area of the cell to be tracked i;

[0071] Wherein d represents the Euclidean distance (x m , y m ) and (x n , y n ) represent the coordinates of the cells to be tracked m and n respectively; Eccent i represents the eccentricity of the i-th cell to be tracked, Eccent i =(4∏Area) / Perimeter 2 , wherein Area represents the cell area, which can be quantitatively described and evaluated by the pixel points occupied by each cell, and Perimeter represents the perimeter of the cell, which can be quantitatively evaluated by the discrete points on the cell boundary.

[0072] Because the definition of the loss function makes the loss function overall smaller, in order to increase the screening effect of the loss function, the reciprocal of the loss function is selected as the judgment basis for tracking, and the reciprocal of the loss function is called the credibility function. The greater the credibility function value is, the greater the credibility of tracking is.

[0073] Among them, the improved ant colony algorithm is used to reflect the cell movement distance, and the distance is converted into the visibility of the ant. The ant only has a moving tendency to the target within its visible range, and has a higher moving probability to the target with a shorter distance.

[0074] Among them, the global movement path avoids the ant colony from falling into local optimization by adjusting the evaporation coefficient, and ignores the selection of the global optimal path. When the current optimal path is obviously shortened, the evaporation coefficient is reduced by the same proportion, the beneficial results of pheromone are maintained, and the optimization is better focused; on the contrary, when the optimal path does not change significantly, the evaporation coefficient is increased, the effect of pheromone is reduced, and the ant colony has more opportunities to choose, expanding the search space of the ant colony.

[0075] The multi-ant colony cell tracking method (MACT) is established, and the specific steps are as follows:

[0076] Step one, in the first frame image, based on the detection result of yolov7, a plurality of ants are distributed near the centroid in each detection box with Gaussian density distribution, and the pheromones of each ant colony are independent of each other and have their own independent numbers.

[0077] Step two, provide the formula:

[0078]

[0079] Among them, τ n is a constant, which is set to 0.01, is the reciprocal of the distance,

[0080] is the initial information of the path ij at the initial time F corresponding to the information at time F.

[0081] According to the above formula, the initial pheromone of each to-be-tracked centroid of the current frame and the next frame is given. The shorter the distance between the cells is, the greater the initial pheromone concentration is, and the ant is more likely to move to the cell with a shorter distance.

[0082] Step three, a greedy algorithm is used to improve the initial search efficiency of the ant colony. The greedy algorithm refers to that the ant colony only focuses on the optimal path in front without considering the global optimal path, and the specific implementation method is:

[0083] Set a common tracking target for all ant colonies in the current frame, each ant is independent of each other, the pheromone of each ant colony is independent of each other, the ant colony can only move to the target within its visible range, when the iteration limit is reached, each ant colony will find the cell it tends to, and the colony with the most ants pointing to the cell is selected as the tracking result.

[0084] Step four, based on the operation of the greedy mode ant colony algorithm, for the cells that failed to track successfully, the ant colony cost function is used for trajectory association, and the cells that still fail to associate successfully are marked as abnormal cells.

[0085] Step five, the movement of the ant colony is based on pheromone and roulette method, the specific formula is as follows:

[0086]

[0087] Where α is the pheromone factor, β is the heuristic factor, both are assigned initial values in the initialization process, and in the specific tracking process, both represent the weight of pheromone and the weight of distance, which are constantly converted into selection probability, then the ant will constantly select the next node according to the formula.

[0088] The parameters are set as follows: the upper limit of the number of iterations is 50, the pheromone evaporation coefficient is 0.15, the pheromone factor is 6, the number of ants is 50, the Q constant is set to 10, and the taboo table is composed of all the center positions of the cells in the next frame.

[0089] The path of the scheme is updated locally between two points

[0090]

[0091] The local pheromone update value for the path between nodes i and j, ρ is the evaporation coefficient, 1-ρ(t) is the residual coefficient, Δτ m(t+1) (i, j) represents the path pheromone increment between nodes i and j, and the value is the reciprocal of the path distance.

[0092] Calculate the path distance value of all schemes, select and record the optimal scheme, and perform global update on the path pheromone of the optimal scheme.

[0093]

[0094] The evaporation coefficient ρ of this global update is obtained by using the above formula.

[0095]

[0096] The optimal path of the scheme is globally updated again using the above formula.

[0097] S5, label the cell which the ant colony tracking algorithm fails to successfully track as an abnormal cell based on the MACT algorithm, and enter an abnormal processing module;

[0098] S51, the ant colony belonging to the cell in the current frame fails to successfully track in the next frame, record the position information of the cell to which the ant colony belongs, and label it as a moving-out or apoptosis abnormal cell;

[0099] S52, the ant colony belonging to the cell in the next frame is not successfully tracked in the current frame, record the position information of the cell to which the ant colony belongs, and label it as a moving-in or division abnormal cell;

[0100] S53, for the moving-out or apoptosis abnormal cell, according to the position information of the abnormal cell, if the position is the boundary of the microscope image, it is determined as a cell moving-out event, and the number, shape, area, eccentricity, last center point position of the moving-out image and other information of the moving-out cell are recorded; if the position is not the boundary of the microscope image, it is determined as a cell apoptosis event, the tracking of the cell is directly ended and the ant colony belonging to the cell is deleted;

[0101] S54, for the moving-in or division abnormal cell, according to the position information of the abnormal cell, if the position is the boundary of the microscope image, it is determined as a cell moving-in event, and if the position is not the boundary of the microscope image, it is determined as a cell division event, the child cells are obtained by comparing with the MACT algorithm, the child cells and the parent cells are re-assigned with numbers and corresponding ant colonies, and the child cells are given an initial number of ants;

[0102] S6, analyze and predict the text information of the centroid coordinates using Kalman filtering, comprehensively consider the speed, acceleration and displacement information in the cell historical data, obtain reasonable prediction values, combine the prediction obtained by Kalman filtering with the position under the given observation obtained by the ant colony algorithm, and obtain the final cell tracking trajectory;

[0103] For the first two frames of cell image sequences, since there is no past value, Kalman filtering cannot be used for prediction, therefore, the detection results obtained by the MACT algorithm are directly used as the output of cell tracking;

[0104] For the kth and subsequent cell image sequences (k >= 3), the cell coordinate difference value of the k-1th and k-2th frames is approximately regarded as the change rate of the k-1th and kth frames, and the Kalman filtering related formula is as follows:

[0105]

[0106]

[0107]

[0108]

[0109]

[0110]

[0111]

[0112]

[0113] P xk =(I-K xk H)P xk-1

[0114] P yk =(I-K yk H)P yk-1

[0115] wherein, is the tracking box x-axis coordinate state variable in the kth frame of cell image sequence, is the tracking box x-axis coordinate in the k-1th frame of cell image sequence, A is the matrix of state transition, ω k-1 is the noise variable, is the tracking box x-axis state prediction covariance matrix in the kth frame of cell image sequence, P xk-1 is the x-axis covariance matrix at the time of tracking box prediction in the k-1th frame of cell image sequence, H is the observation matrix.

[0116] The observation value of the current step is directly obtained by using the detection result of the MACT algorithm, the historical data is integrated and the prediction value of the current step is reasonably predicted by Kalman filtering, and finally the tracking result is obtained by the weighted sum of the prediction value and the observation value based on the Kalman filtering gain coefficient after fully considering the reliability of the observation value and the prediction value.

[0117] S7, the trajectory remodeling function of the ant colony algorithm in MACT is used to avoid generating redundant tracking trajectory fragments and achieve a global optimal solution;

[0118] S71, the trajectory with too small number of ant colonies in the tracking trajectory is removed;

[0119] S72, for the cells in the cell immigration event, the similarity between the cells and the previous emigrated cells is judged, and if the similarity is greater than the immigration and emigration threshold, the immigration cell is identified as the re-immigration of the previous emigrated cell, and the immigration cell is given the previous number and label, thereby enhancing the rigor of the tracking system.

[0120] S73, the trajectory remodeling function of the ant colony algorithm in MACT is used.

[0121] S8, cycle steps S4 to S7, from the initial frame to the last frame image, thus completing the tracking of all cells.

[0122] To sum up, the present technique can effectively liberate productivity, in the face of complex microscopic image sequence under the cell, through the accurate detection of yolov7, the synergistic work of ant colony algorithm and Kalman filter, the reasonable identification of complex cell behavior is realized, and the multi-cell tracking under complex conditions is realized.

[0123] Finally, it should be noted that: the above only for the preferred embodiments of the present application, and not for limiting the present application, although the foregoing detailed description of the present application is made with reference to the foregoing embodiments, for those skilled in the art, it still can be modified to the technical scheme recorded in the foregoing embodiments, or equivalent replacement of some technical features. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A cell tracking technique method based on deep learning and ant colony algorithm, characterized in that: The method comprises the following steps: S1, labelling the cell image sequence under the microscope using labelme, and the label name is the corresponding cell category name, and the storage format is selected as yolo format; S2, the labeled cell image sequence is transmitted to the yolov7 network for training to obtain the optimal target detection weight, and the yolov7 network is modified to make it more suitable for the medical image field and output the center point coordinates of the detection frame; S3, using labelme to label the target detection frame obtained by yolov7, and transmitting the labeled cell segmentation data set to the U-Net network for training to obtain the accurate segmentation image of the cell in each target detection frame, and recording the cell area, cell eccentricity contour information; S4, a cost function of the ant colony algorithm is established, and based on the cell contour information, cell movement distance and global movement path information, a multi-ant colony cell tracking method MACT is established, so that the cell movement trajectory under the given observation is obtained through the ant colony algorithm; The cost function of the ant colony algorithm is established in step S4, and based on the cell contour information, cell movement distance and global movement path information, comprising: The cost function of the ant colony algorithm is established as follows: ; Wherein, I represents the detected cell to be tracked in the current frame, j represents the cell to be tracked in the next frame, and alpha, beta and gamma represent the weights of distance, eccentricity and area respectively, and the sum of the weights is 1; ， ， ; wherein n represents the number of all the cells to be tracked in the next frame, represents the distance between the cell to be tracked i and the cell to be tracked j, represents the eccentricity of the cell to be tracked i, represents the area of the cell to be tracked i; The reciprocal of the loss function is selected as the judgment basis for tracking, which is used to increase the screening effect of the loss function; S5, based on the MACT algorithm, the cells that cannot be successfully tracked by the ant colony tracking algorithm are marked as abnormal cells and enter the abnormal processing module; S6, Kalman filtering is introduced into the MACT algorithm, the moving direction of the cell is predicted through the historical information, and the prediction probability is equivalent to the pheromone of the ant colony algorithm, which is integrated into the subsequent cell tracking, the speed, acceleration and displacement information in the historical data of the cell are comprehensively considered to obtain a reasonable prediction value, the prediction obtained by Kalman filtering is combined with the position under the given observation obtained by the ant colony algorithm to obtain the final cell tracking trajectory; In step S6, the prediction obtained by Kalman filtering is combined with the position under the given observation obtained by the ant colony algorithm, comprising: Using Kalman filtering algorithm, based on the position, speed and acceleration information at past time, the position prediction value of the current frame is obtained; The observation value obtained by the MACT algorithm is set as the observation value; The detection result of the MACT algorithm is used to directly obtain the observation value of the current step, the historical data is integrated and the prediction value of the current step is reasonably predicted through Kalman filtering, and finally the tracking result is obtained by the weighted sum of the prediction value based on the Kalman filtering gain coefficient and the observation value after fully considering the reliability of the observation value and the prediction value; S7, the trajectory remodeling function of the ant colony algorithm in MACT is used to further optimize the existing tracking trajectory; S8, the steps S4 to S7 are repeated from the initial frame to the last frame of image, and thus the tracking of all cells is completed.

2. The cell tracking technique method based on deep learning and ant colony algorithm according to claim 1, characterized in that: In step S5, the cells that cannot be successfully tracked by the ant colony algorithm are marked as abnormal cells, comprising: The cell to which the ant colony in the current frame belongs fails to be successfully tracked in the next frame, and the position information of the cell to which the ant colony belongs is recorded, which is marked as a removed or apoptotic abnormal cell; The cell to which the ant colony in the next frame belongs is not successfully tracked in the current frame, and the position information of the cell to which the ant colony belongs is recorded, which is marked as a removed or apoptotic abnormal cell.

3. The cell tracking technique method based on deep learning and ant colony algorithm according to claim 1, characterized in that: The trajectory remodeling function of the ant colony algorithm in the MACT in the step S7 includes: Eliminate the trajectories with too small number of ant colonies in the tracking trajectory; Design a cell removal and import module to determine the imported cells that have been removed and assign the number and label of the imported cells that have been removed.

Images