The Radiomics Models For Hepatic Tumor Classification Assessment

In recent years,the prevention and treatment of hepatic tumors are more severe than ever and cause a heavy burden on society and families.The preoperative evaluation for hepatic tumors is one of the key factors to improve clinical decision-making as well as the prognosis of patients.Radiomics method based on high-throughput medical imagines can quantitatively describe the heterogeneity inside the tumors which is promising to be a reliable solution for non-invasively preoperative evaluation.In this work,the traditional radiomics method is used as the basic framework while feature engineering and deep-learning feature extraction are applied to enrich the feature set.Moreover,through artificial sample synthesis method and cross-validation,the problem of sample imbalance is solved.Finally,non-linear classification algorithms and convolutional neural networks are implemented to build radiomics models for hepatic tumor classification and lymph node metastasis status assessment.The main work completed is as follows:First,a hepatic tumor classification model based on random forest and multiple logistic regression algorithm is proposed to realize the differential diagnosis of hepatic epithelioid angiomyolipoma.The model can effectively distinguish three types of hepatic tumors under severe sample imbalance which can provide suggestions for the treatment strategies.Second,we develop a lymph node metastasis status classifier according to deeplearning radiomics method.The classifier integrated deep-learning features and radiomics features performs much better than the traditional radiomics model.We also purify the feature set with symbolic regression in order to improve the prediction accuracy and training efficiency.Furthermore,it is confirmed that the status classifier can be fine-tuned to a lymph node metastasis risk stratification classifier with subgroup analysis which proves the generalization ability of proposed classifier.Third,an in-depth research on the clinical application of the models is put forward.Nomograms are plotted by Cox hazard proportional regression.Moreover,a horizontal comparison between the naked-eye assessment and radiomics model prediction shows the deep-learning radiomics model can significantly improve the preventive evaluation.In conclude,the proposed models have considerable prediction performance and clinical application prospect which can realize the accurate assessment of hepatic tumors.The results put forward by this research are expected to be reliable tools to promote the realization of individualized treatment strategies for patients with hepatic tumors.