Research On Deep Learning Algorithm To Analyze Cellular Calcium Signal

Calcium ion(Ca²⁺),as a common second messenger in the cell,participates in almost all important cellular activities.Analyzing the law of calcium ion concentration（[Ca2+]i）over time,called calcium signal,helps to understand cell activity.The existing calcium signal analysis process needs to be manually combined with semi-automatic data processing,and the efficiency is low.Therefore,this article developed an automatic processing method for calcium signals,as follows:（1）Automatic recognition of calcium fluorescence area in a single cellCollect and construct the calcium imaging data set of chondrocytes,BMSCs（Bone Marrow Mesenchymal Stem Cells）and MC3T3-E1（Mouse Embryoosteoblastprecursor Cells）cells,and train FCN-O（Fully Convolutional Networks-Optimized）model through the FCN network（Fully Convolutional Networks）is used for the identification of cartilage cells in the picture.The results show that FCN-O can identify the pixels belonging to the intracellular calcium fluorescence area in the picture with a Dice value of 80%,which improves the efficiency by more than 90%compared with manual work.And using BMSCs and MC3T3 to verify FCN-O,the Dice value is 77%after corresponding fine-tuning,which proves that the FCN-O model has certain robustness.（2）Automatic collection of calcium curveBased on the FCN-O model,a cell tracking algorithm is constructed to classify the FCN-O binary mask of unclassified cells in each frame of the video to each cell.After the classification is completed,obtain the binary mask sequence of each cell in the video over time,find the binary mask when its area is the largest,and calculate the average gray value in turn according to the mask to obtain the calcium curve of each cell.The results show that the Person coefficient between the calcium curve obtained by the cell tracking algorithm and the calcium curve extracted manually is 99.5%.（3）Automatic identification of calcium response peakCollect and establish calcium curve data sets of chondrocytes,BMSCs and MC3T3-E1 cells,establish LSTM-F-O（Long Short-Term Memory With Fully Connected Layer-Optimized）based on the LSTM（Long Short-Term Memory）network,and summarize the manual Labeling experience is implemented by algorithm CPEM（Calcium Peaks Extractiong Model）and CPEM-LSTM-F-O（Calcium Peaks Extracting Model-Long Short-Term Memory With Fully Connected Layer-Optimized）model combined with CPEM and LSTM-F-O models for calcium Identification of the calcium response peak in the curve.The results show that models such as LSTM-F-O,CPEM and CPEM-LSTM-F-O can better identify the calcium response peaks in the calcium curve.On the chondrocyte test set,their F1-scores are 84%,92.6%,and 83.5%,respectively and different models are suitable for different data processing needs.（4）Automatic calculation of biological dataCombined with the automatic cell identification,calcium curve collection and calcium response peak identification methods,the automatic calculation of biological data is realized by algorithm programming.The results show that programming methods can effectively reduce the amount of repetitive work and improve the efficiency of biological data processing.