Cell Image Segmentation Based On Neural Ordinary Differential Equation And Marker Watershed Algorithm

Cell image segmentation is an important technique and a difficult problem in medical image processing.Currently,there is no universally effective segmentation method that can be applied to cell segmentation,mainly because of the varying size of cell images,the complexity of the components and the tendency to overlap between cells.Therefore,cell image segmentation is a difficult and challenging task.Traditional cell image segmentation methods cannot be directly used for accurate segmentation of cells,but with the development of computer technology,deep learning has become increasingly popular in medical image processing.Deep learning is a computational model that resembles the human cognitive system and can be used effectively in different applications.In this paper,we study its application in cell image segmentation,and the main work is as follows.(1)In this paper,a new deep learning network,Attention NODE-UNET++,is proposed for red blood cell image segmentation based on the fully convolution neural network UNet++,which combines neural differential equations(NODE)and Attention Mechanism(AM).The network integrates the advantages of neural ordinary differential equation and attention mechanism.In the encoder stage,a structure with convolution block and neural ordinary differential equation module is designed,which can better extract the shallow features of the image with less parameters.Moreover,the network has a strictly monitored encoder-decoder structure and integrated attention gate jump connection,which can effectively improve the segmentation efficiency of the network and the accuracy of red blood cell image segmentation.The algorithm in this paper achieves better results compared to Attention U-Net,Res Net,and U-Net++ networks.Experiments were done on two datasets separately and the results showed that the Dice coefficient,pixel accuracy,mean pixel accuracy,and mean intersection over union in MISP data set reach 93.47%,96.62%,95.93%and 93.25%,respectively.In the dataset of Cell,the Dice coefficient is 94.69%,the pixel accuracy is 97.43%,the mean pixel accuracy is 96.58%,and the mean intersection over union is 94.42%.(2)In this paper,based on Attention NODE-UNet++ network,combined with Marker Watershed algorithm(MW),overlapping and adhering red blood cells are segmented together.Firstly,the image is adjusted,labeled and data enhanced.Then,the red blood cell image is preliminarily segmented by the Attention NODE-UNet++ network to obtain the probability gray image.Finally,the obtained probabilistic grayscale maps are input to the labeled watershed algorithm for secondary segmentation,which is used to obtain the final segmentation results of red blood cell images.Experiments were done on two datasets separately,and the results showed that the Dice coefficient in MISP data set reaches 95.52%,the pixel accuracy reaches 98.82%,the mean pixel accuracy reaches 97.46%,the mean intersection over union reaches 95.73%,the Dice coefficient in Cell data set reaches 96.87%,the pixel accuracy reaches 99.16%,the mean pixel accuracy reaches 98.53%,and the mean intersection over union reaches 96.58%.Compared with the algorithms in literature [6],literature [16] and literature [18],the effect of extracting single red blood cell is better.