| Many biological images need to be segmented before being analyzed.For example,before dynamic tracking and life prediction of various physiological changes of the common organism caenorhabditis elegans,image segmentation of its microscopic images is required.The image segmentation of colon cancer cells will facilitate the diagnosis of this disease,and the segmentation of mouse cells will also improve the detection accuracy of mouse embryos,so as to facilitate the research of the embryos of this species.The specific work of this thesis is as follows:First,in this thesis,the images of C.elegans,colon cancer cells and mouse embryos were taken as data sets,and the images were trimmed by setting a sliding window of equal step length to make their sizes meet the training requirements.Aiming at the problem of a small number of images in the data set,geometric transformation and pixel transformation were used to expand the data,which improved the diversity and integrity of the data set and the generalization ability of the model.Second,this thesis proposes a method to improve the image segmentation effect under uneven illumination,and improves the generation of confrontation network.Although Unet network can efficiently use the characteristics of different layers and can also be used as a generator of confrontation network,there are still many shortcomings: first,the convolution extraction function of each layer will lead to the loss of many details,thus affecting the segmentation effect.Secondly,the number of network layers is shallow,the number of feature learning is limited,and the expression ability is weak.Therefore,the improved residual splicing link structure can be used as the feature extraction module in this thesis,and it can be generated after being improved in combination with the design idea of Unet.The attention mechanism can also be added in the early stage of training.At the same time,combined with the model compression technology,the irrelevant weights of the network can be eliminated and the network size can be compressed.This effectively solves the problem of feature loss,reduces network parameters,and largely prevents the gradient from disappearing and exploding,thus improving the feature extraction performance of the network and realizing efficient and fast training.Third,in view of the fact that the data set of Caenorhabditis elegans has uneven illumination and is difficult to preprocess,this thesis uses the bottom hat transform to deal with the impact of uneven illumination on the data of Caenorhabditis elegans elegans data set for image segmentation,and then uses the improved generation antagonism network model to segment the image of Caenorhabditis elegans,and makes a number of control experiments on it,such as Otsu method,Unet method and cellprofile method,wait.Finally,it is found that the improved generation antagonism network is superior to other methods in many image segmentation accuracy indicators.Then the improved generation antagonism network is used to test the data set of colon cancer cell image and mouse embryonic cell image,and the result of image segmentation accuracy is also good. |
