Nuclear Segmentation Based On Deep Convolut Ion Neural Network

Deep learning is one of the latest trends in machine learning and artificial intelligence.Nuclear segmentation based on deep neural network is an important branch of computer vision.It has important research significance and wide application value in drug development,disease diagnosis and other fields.Nuclear segmentation based on deep convolution neural network mainly includes feature extraction,recognition,segmentation prediction and other functional modules,among which extracting discriminant features is the key to affect the accuracy of nuclear segmentation.Generally,networks with good generalization ability and high discriminant characteristics need a certain scale of a high-quality training set.However,the artificially labelled cell nucleus segmentation image is very precious.How to obtain good model generalization ability and better segmentation accuracy on a small-scale of data sets has become the main research direction of this thesis.In addition,the decisive factors affecting the ability of feature extraction of deep network come from two aspects:network architecture design and loss function design.The improvement of network structure and loss function will greatly affect the quality of nuclear segmentation.In addition,because the image itself has the characteristics of uneven brightness distribution,different number and size of nuclei,adhesion and stacking between cells,it will also lead to the abnormal difficulty of nuclear segmentation.Based on the above analysis,in order to achieve a cell nucleus segmentation model with good generalization ability and better segmentation accuracy on a small data sets,the main research work of this thesis is as follows:(1)Based on U-net full convolution neural network,a nuclear image segmentation algorithm is proposed in this thesis.Firstly,in order to improve the generalization and recognition ability of the network,data augmentation is necessary.Then K-means clustering is done according to HSV colour space on the data sets,and different clustering is selected for cross-validation in U-net network training.Secondly,in view of cell adhesion,this thesis uses corrosion operation on the groud truth mask in pre-process to ensure that each nucleus is separated and then uses expansion operation to adjust the final results.Finally,the loss function of U-net network is improved to obtain better segmentation results.Experiments show that the proposed method is better than traditional segmentation algorithms,such as Ostu and watershed algorithm.Based on the quantitative analysis of nucleus segmentation results,the PA on Broad Bioimage Benchmark Collection 038 data set reached 94.944%,the mIoU reached 0.55452.(2)In order to make the U-net network deal with the situation of adhesion cells better,using the thought of multi-task learning,changing the input and output of the U-net network during training,making the network pay attention to the nucleus,the boundaries of the adhered nucleus and the background at the same time.Secondly,in order to make the segmentation model have better generalization ability,the feature extraction part of the U-net network is replaced by the first layers of the pre-training VGG16 model using the method of transfer learning.The network is initialized with VGG16 model which converges and generalizes on ImageNet,and then fine-tuned on small data' sets to make the network adapt to the problem of cell nucleus segmentation.Finally,in the post-processing stage,the segmentation results are analyzed,and a series of morphological operations are used to refine the final segmentation results.Quantitative analysis and horizontal comparison of the algorithms in this paper prove the efficiency and robustness of the algorithm.The PA on Broad Bioimage Benchmark Collection 038 data set is 95.437%,the mIoU reached 0.60241.