Author Affiliations
1Sino-German College of Intelligent Manufacturing, Shenzhen Technology University, Shenzhen 518118, Guangdong , China2College of Health Science and Enviroment Engineering, Shenzhen Technology University, Shenzhen 518118, Guangdong , China3Key Laboratory of Advanced Optical Precision Manufacturing Technology of Guangdong Provincial Higher Education Institute, Shenzhen 518118, Guangdong , Chinashow less
Fig. 1. Angularly multiplexed OPT system
Fig. 2. Reconstructed cell images. (a) XZ direction; (b) YZ direction
Fig. 3. Flow chart of zebrafish cell identification and tracking model
Fig. 4. Structure diagram of zebrafish cell recognition and tracking model
Fig. 5. Three-dimensional cell tracking framework
Fig. 6. Three-dimensional tracking algorithm flow chart
Fig. 7. Mask generated by Labelme. (a) Original map of cells; (b) mask map
Fig. 8. Data augmentation (a) Original; (b) distortion; (c) overturn; (d) add noise
Fig. 9. Comparison of loss and mAP under different learning rates. (a) Training loss under different learning rates; (b) validation loss under different learning rates; (c) mAP under different learning rates
Fig. 10. mAP change curves under different improvement losses
Fig. 11. Loss of 606 training sets when learning rate is 0.001
Fig. 12. Local comparison of segmentation results. (a) Original cell image; (b) watershed segmentation based on gradient transformation; (c) morphological segmentation; (d) segmentation based on U-Net; (e) segmentation based on Mask R-CNN; (f) segmentation based on Mask R-CNN++
Fig. 13. Effect of segmentation in data augmentation. (a) Distortion; (b) add noise; (c) overturn
Fig. 14. Results of tracking by DeepSort. (a) Part of cells in 20th frame; (b) part of cells in 30th frame
Fig. 15. Cell trajectory (a) Three-dimentional trajectory map of cells; (b) position of cell relative to Z axis of first frame changes
Parameter | Meaning | Value |
---|
STEPS_PER_EPOCH | The number of pictures per epoch in training | 100 | VALIDATION_STEPS | The number of pictures per epoch in validation | 33 | DATASET_TRAIN | The number of training set | 606 | LEARNING_RATE | Determine whether the objective function can converge | 0.001-0.01 | WEIGHT_DECAY | Reduce model over fitting | 0.0001 | IoU_THRESHOLDS | The thresholds of intersection over union | 0.5/0.75 | EPOCH | Number of iterative training | 80 |
|
Table 1. Parameter setting of network
Learning rate | Training time /s | Frequency /(frame·s-1) | mAP11 | mAP21 | mAP60 |
---|
0.01 | 26295 | 0.304 | 0.8926 | 0.6681 | 0.9586 | 0.001 | 26134 | 0.306 | 0.9092 | 0.9205 | 0.9532 | 0.0001 | 26459 | 0.302 | 0.6727 | 0.8138 | 0.8811 |
|
Table 2. Time efficiency and part of the model mAP under different learning rates
Algorithm | XZ plane | YZ plane |
---|
Precision | Recall | Precision | Recall |
---|
Morphology[8] | 71.64 | 56.81 | 84.89 | 71.87 | Watershed[26-27] | 85.28 | 82.73 | 85.24 | 73.31 | U-Net[15] | 97.76 | 68.29 | 97.61 | 67.20 | Mask R-CNN[16] | 96.25 | 95.90 | 94.08 | 96.30 | Mask R-CNN++ | 98.99 | 96.74 | 97.86 | 98.07 |
|
Table 3. Precision and recall of cell segmentation by different algorithms