Research On Artificial Intelligence Of Cancer Screening And Diagnosis Based On Neural Networks

Cancer has become a major threat to human health.Early cancer screen-ing and diagnosis is an effective way to reduce cancer mortality.However,the deteriorating situation of cancer and the relative shortage of professionals are a sharp contradiction that people face in the current cancer situation.The in-telligent screening and diagnosis of cancer provides an effective way to solve this contradiction.In recent years,artificial neural network algorithms,especial-ly deep neural network algorithms,have achieved remarkable results in various fields of artificial intelligence.This data-driven machine learning algorithm re-quires learning from a large amount of data.With the deepening of medical informatization,the amount of data in the medical industry has grown rapid-ly.The rapid accumulation of medical data and the continuous development of artificial neural networks provide powerful materials and tools for intelligent screening and diagnosis of cancer.This thesis focuses on the intelligent screening and diagnosis of cancer.Based on the in-depth study of the characteristics of medical data and the ex-tensive analysis of existing related research work,a series of neural network models for cancer screening and diagnosis are proposed,and performance eval-uation and verification are performed on several public data sets.More specifi-cally,the main contributions of the thesis are presented as follows.Multi-view convolutional neural networks(MV-CNN)for pulmonary CT pulmonary nodules intelligent diagnosis are proposed.CT images have two characteristics.The first one is that the lesion makes up a small part of the entire image.The lesion affects the surrounding tissue,so the surrounding tissue environment of the lesion can also provide useful information for the di-agnosis of medical images.In view of this,a multi-view convolution neural network is proposed.The model uses multiple channels in the input layer and each channel corresponds to a field of view.In this thesis,the model is used to conduct a binary classification(benign and malignant)and a ternary classifica-tion(benign,primary malignant and metastatic malignant)experiments on the LIDC dataset.The error rate for the binary classification is 5.41%and 13.91%for the ternary classification.The results for this dataset are better than the re-sults of other research.The second characteristic of medical images is often composed of many slices that connected to depict the tridimensional structure of the internal organs.In response to this characteristic,the above model is im proved,and a 3D multi-view convolution neural network is proposed.That is to say,the data and network frameworks are organized in 3D mode,which al-lows the convolution to slide in three directions so as to extract features in three dimensions of pulmonary nodules.Different from other related studies,the net-work architecture adopts the 3D variant of the Inception module.On the same data set,the model reduces the error rate to 4.59%for the binary classification and 7.70%for the ternary classification.Fully connected layer first convolutional neural networks(FCLF-CNNs)are proposed for the diagnosis of breast cancer based on cytological features,which are suitable for improving the classification performance of structured data.Fully connected layers are added in front of the convolu-tional layer,which serves to convert the structured data into data representation with better local structure.In practice,the model imposes two loss functions on the fully connected layer before the convolutional layer,and each corresponds to one network architecture.One is the softmax loss function,corresponding to the 1D FCLF-CNN architecture and the other one is the MSE loss function,corresponding to the 2D FCLF-CNN architecture.Then,each framework uses two training methods,including simultaneous training and step-wise training.Therefore,there are four different FCLF-CNN models.The integration of these four networks showed that the resulting ensemble model achieved better results than the other models.The cross validation of this model achieves an accuracy of 98.71%,a sensitivity of 97.60%and a specificity of 99.43%for WDBC,and an accuracy of 99.28%,a specificity of 98.65%and a sensitivity of 99.57%for WBCD.The results for both datasets are competitive compared to the results of other research.Using the proposed FCLF-CNN for asymptomatic cancer screening,based on blood routine,urine routine,and tumor markers as features.And a new method combining a hard classifier and a soft classifier was designed based on extremely unbalanced scenes to output screening result-s.Traditional cancer screening uses relatively balanced data for research and uses only hard classifiers.Asymptoma-based cancer screening is faced with ex-tremely unbalanced scenarios.Through theoretical and experimental methods,it is proved that hard classifiers and soft classifiers are affected by the.degree of class imbalance.Therefore,based on extremely unbalanced scenes,a new method combining a hard classifier and a soft classifier was designed to out-put screening results.For the hard classifier,the maximum mutual information principle is used for threshold selection.For the soft classifier,a strategy for simultaneously outputting the false alarm rate,the false negative rate,and the PPV corresponding to the score is adopted.FCLF-CNN was used and applied to the screening of lung cancer,liver cancer,breast cancer,cervical cancer,and kidney cancer.Compared with other related studies,the models of lung cancer and liver cancer achieved more competitive screening performance,with AUC of 0.8887 and 0.9432,respectively.