Segmentation Of Cell Microscopic Image Based On CNN-CRF

Currently,understanding the structural functions of human brain and behavioral cognition from the level of cells and networks is an important research content of system neuroscience.Considering the large similarity between the brain and the human brain in the structure and gene sequence,accurate and efficient segmentation of rat brain cells is very important for subsequent researches such as recognition,counting,tracking and morphological analysis of brain cells.At present,the image of mouse brain cells photographed by optical microscope often has the characteristics of large amount of data and high tightness.The traditional method of manually segmenting mouse brain cells not only has a large workload and takes a lot of time,but also tends to reduce the precision due to the operational fatigue.With the development of computer vision technology and deep learning technology,automatic cell segmentation of mouse brain slice microscopic image has gradually become an important research content in the field of system neuroscience.In this paper,a segmentation model based on convolutional neural network is designed for the automatic segmentation of mouse brain microscopic images witch has more noise,lower image contrast and larger data.The model expands the neural network receptive field by using a convolution block containing a multi-layer small-core filter,and also improves the cross-entropy function to solve the difficult-to-classify sample problem.A variety of anti-over-fitting techniques are also incorporated to improve the stability,accuracy and generalization of the model.In addition,in order to overcome the problem that CNN model's inverted pyramid down-sampling process leads to excessive refraction and ignoring local information,we proposed a CNN-CRF model by combining conditional random field which can characterize the shape and color features.The model reprocesses the CNN pre-segmented image,and constructs a pairwise relational potential function for each pixel to add global context information,which adds the constraint on spatial and edge information and improves the segmentation precision of the mouse brain cells greatly.We also have transfered the CNN-CRF model to get a better segmentation of both mouse stem cells and Hella cells.The improved CNN-CRF model can effectively extract feature of mouse brain shape and edge at the pixel level,and have better generalization ability than FCN8,Seg Net,random forest and etc.The method effectively realizes the segmentation accuracy of rat brain cells,which is very important for the research and analysis of brain's structure and function.