Research On Nuclear Classification Of Colon Cancer Images Based On Convolutional Neural Networks

Cancer is the main cause of human death in the world,in which the number of colon cancer patients is large and the mortality is high.It has become the third most dangerous cancer affecting human life,but its early diagnosis survival rate is only 50%.Therefore,accurate early diagnosis and analysis of colon cancer has important research value and significance.In recent years,there have been some research results in the early diagnosis of colon cancer by automatic analysis of the image nuclei,but there are some problems such as large feature error of manual extraction and low accuracy of diagnosis.With the continuous development of deep learning,convolutional neural networks can directly extract features from the original image,thus making a major breakthrough in medical image processing.However,because of the time-consuming training and the high heterogeneity of colon cancer image nuclei,it is difficult to classify colon cancer image nuclei efficiently by some previous methods.In order to solve the above problems,this paper firstly studies the influence of full connection layer on the performance of convolutional neural networks and the potential relationship between data set and network from the internal structure of convolutional neural networks,proposes convolution model of overlapping clustering to improve the training speed of network,and refers to the experimental setting of full connection layer on the performance of convolutional neural networks,proposes A New Nuclear Classification Method for Colon Cancer Image based on Clustering Convolution.The main research contents of the paper are as follows:(1)The influence of full connection layer on the performance of convolutional neural networks is studied.This paper proposes three initial network architectures from shallow to deep,namely Net-1,Net-2 and Net-3.The influence of neurons on the performance of convolutional neural networks is studied by gradually increasing the number of neurons in the full connection layer.Finally,the potential relationship between dataset and convolutional neural networks is analyzed by experiments.The experimental results on four datasets CRCHistoPhenotypes,Cifar-10,Cifar-100 and Tiny ImageNet show that the deep layer network does not need more neurons and layers in the full connection layer,but the shallow layer network.For datasets,the deep dataset needs deeper network training.Moreover,when the number of neurons in the fully connected layer of Net-3 was updated to 128 and 4,a classification accuracy of 84.79% was obtained on CRCHistoPhenotypes.(2)A New Nuclear Classification Method for Colon Cancer Image based on Clustering Convolution is proposed.After referring to the experimental setup of the influence of full connection layer on the performance of convolutional neural networks,this paper takes Net-3 network which has the best performance of nuclear classification of colon cancer image as the reference network.Then,combined with the Restricted Boltzmann Machine model and the k-harmonic means and overlapping k-means algorithms(KHM-OKM)overlapping clustering learning method,a convolution model of overlapping clustering is proposed and applied to the preprocessing of colon cancer image nuclear classification model to improve the training time of convolution neural networks.Finally,combined with the new dropout model and residual model,A New Nuclear Classification Method for Colon Cancer Image based on Clustering Convolution is proposed.This method not only reduces the time of network training,but also achieves 85.98% classification accuracy on CRCHistoPhenotypes.In order to verify the generalization ability of this method,we also do experiments and analysis on Cifar-10,Cifar-100 and Tiny ImageNet.