چكيده لاتين
Cellular heterogeneity, even among identical cells, results in differences in their characteristics, necessitating the analysis of each cell individually for more accurate examination. However, isolating single cells from a cell population using current laboratory methods presents significant complexities due to the risk of cell damage. droplet microfluidic technology allows for the encapsulation of cells, including single cells, in separate droplets without causing harm. However, sedimentation and aggregation of cells in microfluidic channels reduce encapsulation efficiency. Consequently, the use of artificial intelligence and deep learning for real-time control of encapsulation conditions is crucial. However, producing sufficient data to train deep learning models is challenging due to clogging in microfluidic channels and the rarity of certain cell types.
This research aims to enhance the quantity of training data by utilizing a Generative Adversarial Network (GAN) to produce diverse and high-quality synthetic data for the automation of droplet microfluidic systems. Two important approaches were followed. In the first approach, synthetic data were generated using a GAN. Subsequently, by combining these data with real data, a training set with 67% synthetic data was created, which was then used to train the YOLOv8s model for the automatic detection and classification of microfluidic droplets in images. Finally, the performance of the resulting model was compared with that of a model trained on images generated using conventional data augmentation methods. In the second approach, a training set with 96% synthetic data was used to train the YOLOv8s model. Analysis of the results from the first approach revealed that the model trained on a combination of real and synthetic data achieved higher performance with an mAP0.5 (mean Average Precision) of 98% compared to the model trained on data generated using conventional data augmentation methods. Additionally, the training process was faster and more stable. Furthermore, the powerful YOLOv8 network achieved a very high processing speed, detecting around 2,338 droplets per second, significantly outperforming previous YOLO versions for both models. In the second approach, the resulting model achieved an mAP0.5 of 92%, demonstrating that the use of a GAN to compensate for severe data scarcity was highly successful.Keywords: Autometrics algorithm, Adaptive lasso regression, Lasso regression, Least trimmed squares, Least median of squares, Lesat squares method, Ridge regression.
