Cell Image Detection And Segmentation Based On Deep Learning

Cell detection and cell segmentation are common tasks in biological-image analysis.Due to differences in imaging equipment,cell types,morphology and density,imaging analysis is a big burden for most scientific research because it often require highly customized algorithms,even tedious manual operations.In the past few years,deep learning has been widely applied in almost all image processing tasks due to its huge advantages in automatic feature extraction,good generalization,and high precision.In several competitions of biological image processing,deep learning-based methods surpassed conventional methods by a big margin.However,the adaptation of these methods in practical research still needs nonnegligible efforts for customization that requires solid background on the algorithm details.In this thesis,a computational framework is designed for the common tasks of 2D cell image detection,2D cell image segmentation and 3D cell image segmentation,and a corresponding set of algorithms is provided,and the framework is applied to practical experimental data analysis tasks.1.Apply YOLOv5 to the detection task of two-dimensional cell images,and use a dataset with rich cell types to train the model to increase the generalization of the model.Subsequently,the metrics of YOLOv5 s and YOLOv5 m on the test set were tested,and the indicators of the two at the lower Intersection-over-Union threshold were very close.Then this thesis tests the accuracy of the faster YOLOv5 s model of the two after training with a single category of cell image datasets of different sizes.The results show that the model can achieve almost the same performance as all 45 training sets on the training set composed of 15 images,which to a certain extent illustrates the feasibility of the framework,that is,the pre-trained model can be applied to cell image analysis tasks by fine-tuning with a small amount of training data.2.The algorithm based on U-Net is applied to the instance segmentation task of two-dimensional cell images,and the U-Net algorithm is improved by adding gradient vector fields.The results on the test set show that the improved instance segmentation accuracy of the algorithm is significantly better than the original U-Net.The resulting model was then applied to the data processing of macrophage experiments,and the algorithm achieved satisfactory results without fine-tuning.3.According to the 3D imaging data such as mouse brain and monkey brain obtained by the 3D imaging equipment developed by our research group,the 3D cell instance segmentation algorithm based on U-Net is improved.The algorithm combines the two-dimensional segmentation results in three directions of the three-dimensional image to obtain three-dimensional results,which reduces the requirements for training data and computing resources.For the problem that the scale of 3D data is too large,this paper effectively improves the processing speed of the algorithm by using the reparameter module,customized network structure parameters,pipeline and parallelization,and the algorithm throughput can be increased by 2 to 12 times.The algorithm was applied to the analysis of 3D imaging data of the whole mouse brain,and achieved satisfactory results.