Methodology Research Of Nucleus Image Segmentation Based On Deep Learning

With the rapid development and application of Medical Imaging Technology these years,it has been used to improve the efficiency of doctors' diagnosis and reduce the misdiagnosis rate on more and more medical diagnostics.In this case,medical institutions need to deal with more and more medical images,segmenting medical images is crucial among them.In previous clinical diagnoses,experienced doctors usually segmented manually,which was time-consuming and energy-consuming and the results are unstable and divergent.Therefore,it is very important to segment medical images accurately,quickly and efficiently.Traditional image segmentation methods generally need to extract some features manually,which usually contain low-level image features,such as the edges,corners,textures and lines of the image.These low-level image features tend to be less robust and susceptible to environmental impacts.Some researches show that in general,high-level and abstract features are more important for image segmentation.And deep learning networks can start from the original pixel-wise features,through the layer of convolution operation,and finally extract the high level features which are crucial to the image segmentation,and then use these features to segment the image.In this paper,I propose a Fully Convolutional Network(FCN),U-net and improved U-net to segment the nucleus,which change the fully-connected layers in the traditional convolution neural network to the convolution layers,and use the deconvolution structure to implement the upsampling.Compared to FCN adding the features by point,U-net splices the feature map on the channel dimension to realize the fusion of features,and uses the skip structure to combine the low-level and the high-level features.Compared with FCN,U-net has achieved better segmentation results,but there still exist some defects such as poor segmentation of small nuclei,rough edges,under segmentation and over segmentation.Therefore,the last chapter of this paper not only uses a more complex convolution neural network to construct the encoder,initializes using the pre-training parameters,and adds dilated convolutions to the encoder,as well as improves the loss function of the original Unet.At the end of the paper,we analyze the segmentation results of FCN,U-net and improved U-net segmenting nucleus.It shows that,with the improvement from FCN to U-net to improved U-net,the segmentation effect of nucleus images is getting better and better,robustness and robustness are steadily enhanced,the segmentation ability of edge details and smaller nuclei is more and more strong,and the phenomenon of over segmentation and under division is gradually reduced.Which proves that the improvement of FCN and U-net has achieved the corresponding results.