Research On The White Blood Cell Nuclear And Cytoplasm Segmentation Based On Convolutional Neural Network

Cell image segmentation has been an important research object in the field of medical image.In the field of white blood cell image segmentation,the variety of white blood cell types resulting in different cell morphology,increased the difficulty of segmentation.At the same time,there are a large number of red blood cell adhere with white blood cell in blood smears,and the difference in imaging results caused by different staining techniques and imaging methods are also difficult problems in white blood cell segmentation.The research purpose of this thesis is to use computer technology to replace the manual repetitive work.This thesis uses the strong self-learning ability of deep learning to solve difficult problems in the field of cell segmentation,and improve the accuracy and efficiency of cell segmentation.According to the above question this thesis has done the following works:(1)Based on the Convolutional Neural Network,this thesis combines the structural advantages of UNet and EfficientNet to create a new white blood cell nuclear segmentation model.This model combines the UNet structure with the EfficientNet structure,and uses the pre-trained EfficientNet optimal structure to extract features to improve the ability of feature learning.At the same time,the fusion of multi-scale features in the network is achieved through skip connections,which improves the segmentation accuracy of the model.A Depthwise Convolution is also applied to the model to improve the training speed of the model.(2)The model is tested on three different data sets,and compared with other methods such as UNet,SegNet and so on.In the experimental process,the used data were firstly preprocessed by resizing,grayscale immobilization and data augmentation,and then trained on three data sets.The experiments verify that the model in this thesis can achieve clear nuclear and cytoplasm edge under different segmentation environments and keep the stable state in the segmentation effect.At the same time,compared with other segmentation methods,the segmentation effect and indexes of the model in this thesis are better than those of other methods on three data sets,among which F1 value reaches 0.963,0.977,and 0.866,respectively.(3)In this thesis,the morphological information obtained from data set 3 and test set was analyzed.I obtained the information of the five types of white blood cells in nuclear area,cytoplasm area,nucleus-somatic ratio,nucleus-cytoplasm ratio and number of leaves.In addition,the K-nearest neighbor algorithm was used to achieve the classification of white blood cells based on the obtained morphological information.The classification accuracy in the data set 3 test set can reach 93.3%.