Prediction Of Survival Rate Of Head And Neck Cancer Based On Convolutional Neural Network

Head and neck cancer has many primary lesions and has a high fatality rate,ranking it among the top tumors of the whole body.Survival prediction for cancer patients can identify potentially high-risk patients in order to determine the treatment measures as early as possible to prevent the disease from worsening.Convolutional neural networks have shown great advantages in the field of image processing and analysis,and have been applied in classification,diagnostic,separation and so on.Therefore,this paper is based on the CNN method,and predicts the survival rate according to the medical images of patients with H&N cancer.The purpose is to provide a consultation for clinical treatment and promote the development of individualized precision medicine.Transfer learning method realizes classification prediction by calculating the correlation between the source domain and the target domain,and its experimental data sets are mostly natural images.However,medical images are mostly for specific problems(such as the heart,breasts and other different organs),which are not universal compared to natural images.The value of transfer learning for medical image classification and prediction is very limited.Therefore,this paper build an end-to-end deep learning network model based on the PET and CT images of patients with head and neck cancer to predict the survival rate of patients,instead of using transfer learning or auxiliary machine learning algorithms.Kernel Principle Component Analysis(KPCA)is used to initialize the convolution kernel.Because KPCA cannot find the optimal main feature direction and retains the traditional PCA’s sensitivity to noise,this paper proposes a method to improve the structure of the KPCA algorithm.For individual differences of H&N cancer tumors,we refer to existing research models to study the texture features that affect the survival rate of patients with H&N cancer.The data set comes from The Cancer Imaging Archive,with a total of 298 patients.56 patients(approximately 19%)died during the follow-up period.Extract the center slice of the tumor ROI area,Do wavelet transform texture enhancement experiments on CT images,construct different CNN prediction models for FDG-PET and CT images,evaluate the performance of the model through the 3-fold cross-validation method,and build a dense net model combining BAM or CBAM for comparative experiments.Finally,using the GLDM to extract texture features from the patient’s CT images,Kaplan-Meier is used to make the relationship curve between texture feature parameters and survival time,and the log-rank test verifies its statistics.Experimental results show that the accuracy of the CNN prediction model using CT images as a data set is better than FDG-PET and wavelet transform texture enhanced CT.AUC,sensitivity,and specificity are: FDG-PET(55.6%,92.9%,18.5%),CT(70.5%,67.4%,76.5%),wavelet transform texture enhancement CT(60.9%,51.7%,77.8%).The improved KPCA algorithm has a significant improvement effect on the prediction performance of CNN.Using the texture features extracted by the GLDM,it is found that the texture parameter "Energy" has significant statistical difference from the patient’s survival time.The research results in this paper show that deep learning has the ability and potential to predict the survival rate of medical images before treatment of head and neck cancer,and can provide guidance information for clinical pathological diagnosis,and to a certain extent contribute to the realization of accurate diagnosis and treatment of patients with head and neck cancer.