Research On Recognition Of Tumor Cell Images Based On Convolutional Neural Network

Cancer is currently the second most deadly disease in the world.It is a type of disease caused by abnormal cell proliferation in the human body.Because the new or old cells of the human body cannot be processed normally,the growth and spread of abnormal cells is caused.In the medical field,the diagnosis of cancer is mainly a doctor's manual analysis and diagnosis through pathological sections,and the results are obtained.This method is particularly time-consuming and difficult to diagnose.Computer-assisted methods have been developed to analyze pathological sections.However,the traditional method is mainly for relatively low-level image analysis tasks(such as color normalization,kernel segmentation,and feature extraction).So it cannot be applied clinically.With the large-scale implementation and promotion of artificial intelligence in different fields,deep learning based on convolutional neural networks has become a focus and hotspot in the field of scientific research in recent years.Deep learning has been researched and applied in the medical field.This article applies deep learning technology to the classification of medical tumor cell images to realize the diagnosis of cancer diseases by computer-assisted pathologists.This paper proposes a system for tumor cell classification based on convolutional neural networks.It is mainly used to detect cancer metastasis on histological images.The research mainly includes the following stages: 1.Use image processing technology to detect the region of interest(ROI);2.Construct a data set,extract positive and negative small image blocks from the ROI region;3.Use convolution The neural network classifies the small image blocks;4.Establishes the tumor probability heat map;5.Post-processing of the whole pathology slide image(Whole-Slide Images,WSI)heat map.The convolutional neural network framework adopted in this paper is based on the Inception-V3 network structure.In order to improve the recognition accuracy of the classifier model and the comprehensive performance score of the classifier,itattempts to modify the original network framework.We made two improvements to the Inception-V3 network framework structure.The first modification of the network is to add a small-sized convolutional layer and pooling layer before the fully connected layer;the second modification is to use the parallel branch sampling method between the sixth and seventh layers of the network.And after many optimizations,a new network architecture is obtained.The experimental data comes from the official data released by the Camelyon16 Challenge,and uses image processing technology to build its own training data set.Compared with the baseline model(Baseline),the accuracy of the trained model has been improved.The area of the improved model under the receiver working curve has reached 90.61%(AUC =0.9061),which is nearly 5% higher than Baseline,and the FROC score(0.7784)has also been increased by about 2%.This result can assist pathologists to improve the accuracy of pathological diagnosis.