Deep Learning-based PET/CT Radiomics Study Of Non-small Cell Lung Cancer

Lung cancer is one of the most common malignant tumors in the world.Its incidence and mortality are increasing year by year,posing a serious threat to human life and health.Therefore,early detection,early diagnosis,correct staging and reasonable treatment of lung cancer are of great significance to prolong the life of lung cancer patients and improve their quality of life.Objective: PET/CT examination is an effective auxiliary method for lung cancer diagnosis,and Ki-67(marker of proliferation Ki-67)is an important marker for predicting tumor cell proliferation and prognosis in patients.Therefore,DenseNet convolutional neural network was adopted in this paper to construct a prediction model for expression level of Ki-67 and survival risk prediction model for non-small cell lung cancer using PET/CT images.Methods: The region of interest(ROI)of PET/CT images of159 patients were delineated and extracted,and the ROI images were equalized by limited contrast histogram,windozed and standardized by pixel value.On this basis,PET and CT images at the same position are superimposed into three-channel combined images to obtain three image datasets of PET image,CT image and combined image,and the image is rotated,flipped,translated and scaled to expand the datasets.Non-small cell lung cancer patients were divided into high expression group and low expression group according to the level of Ki-67 expression.The above three data sets were input into DenseNet convolutional neural network using group label,and the models were trained by ab initio training strategy and transfer learning strategy respectively,to obtain six prediction models of Ki-67 expression level.The score of the model with the best performance was selected as the risk score of patients with Nonsmall cell lung cancer,and the patients were divided into risk grades based on 95% sensitivity threshold,95% specificity threshold and maximum approximation index threshold,and the survival risk assessment model was constructed.Results: Three types of images were trained by ab initio training strategy and transfer learning strategy,and six different models were obtained.Among the six models,the performance of the combined model is better than that of the other two models.PET model has better performance than CT model.The model performance of ab initio training strategy is better than that of transfer learning strategy.Among them,the combination model with ab initio strategy training has the best AUC value of 0.891.Patients with Non-small cell lung cancer were graded with the output score of the optimal model,and there were significant stratification in the overall survival of patients with different risk grades in the test set.Conclusion: The prediction model of Ki-67 expression level based on DenseNet convolutional neural network can effectively predict the level of Ki-67 expression in NSCLC.The deep learning model score of Ki-67 expression level prediction was significantly associated with the prognosis of patients with NSCLC,and was an important prognostic factor of patients.