Application Research Of Cancer Pathology Image Recognition Based On Convolutional Neural Network

Cancer is a very fatal disease,its early symptoms are not obvious,the diagnosis is difficult,and the number of patients increases faster than the training speed of professional scholars.Therefore,the use of computers for medical image-assisted diagnosis and automatic recognition of cancer pathological images have become the expectations of the current industry.This paper makes a classification and recognition study based on deep learning for pathological images in the field of medical imaging.The main work includes:1)Research on prostate cancer pathological image recognition based on convolutional neural network.Design a deep learning-based training program for detecting cancer metastasis from histological images.It includes five stages: detection of regions of interest(ROI,Region of Interest);block sampling and labeling based on ROI information as training data;input the labeled sample blocks into the selected training network model to train neural networks based on block classification.Build a probabilistic heat map of the entire pathological image based on the binary classification results of the sampling blocks in the entire pathological image;post-process the heat map by selecting 5 features to diagnose the tissue section.Use three models of VGG,Inception-v3 and Res Net to train the neural network.The optimal model Inception-v3 achieved a performance of 0.860 under the ROC(receiver operating characteristics)curve under this method,and the performance reached a high level.2)Research on prostate cancer pathological image recognition based on transfer learning.In order to solve the problem of the shortage of prostate cancer pathological images in the first part of the study,it is proposed to use the existing two data sets of retinal eye angle images and breast cancer pathological images to perform transfer learning on 120 cases of prostate cancer pathological images in this paper,and train block-based prostate cancer images Classification model.First,pre-train the two pre-training data sets with VGG,Res Net-50 and Inception-v3 network models.After the model converges,save the model weights,and then enter the prostate cancer pathological image sampling block into the network model to fine-tune the modelparameters.The experimental results show that for the training of prostate cancer pathological images,the effect of transfer learning on the selected pre-training data set is better than that in the first part of the study without transfer learning,and it is better than traditional methods for transfer learning on Image Net Case.3)Research on breast cancer pathological image recognition based on weakly supervised learning.In order to solve the time-consuming and laborious problem of the secondary labeling process in the first two parts of the study,a new method for training the pixel resolution segmentation model of the entire image in the weakly supervised setting for the pathological image of breast cancer is given.The model can be trained under noisy labeled data.First,by focusing on regions with high confidence in the model,the model's learned experience is used to randomly and dynamically sample.Secondly,using the extended KL(Kullback – Leibler divergence)divergence,the extension is reliable for noise labels.The results were verified under the CAMELYON 16 dataset,and the final area under ROC was 0.916.Compared with the traditional fully supervised learning method,this method sacrifices about 3.37% accuracy and obtains about 30.6% improvement in research time efficiency.