Study Of Radiomics On Preoperative Noninvasive Prediction Of Tumor Prognostic Factors

Cancer has seriously affected the improvement of life expectancy,and the number of morbidity and mortality is still increasing.Accurate preoperative diagnosis and prognosis prediction can help doctors in individualized decision-making with important clinical significance.In clinical practice,the heterogeneity of the tumor may lead to bias of preoperative biopsy,and only pathological analysis of whole tumor samples can obtain accurate diagnosis.However,pathological analysis is lagging behind,which does not meet the requirements of accurate preoperative diagnosis.Therefore,accurate preoperative diagnosis and prognosis prediction is a great challenge in clinical work.With the development of computer and medical imaging technology,"Radiomics"that based on big data in medical image emerged at a historic moment.Radiomics aims to transform the heterogeneity information of the tumor contained in medical images into mineable quantitative features.According to the clinical problems,statistical learning and machine learning technologies are used to screen important radiomic features and build prediction models,so as to provide help for doctors'decision-making.It has been widely used in the preoperative noninvasive diagnosis and prognosis prediction of various cancers.In this paper,CT and MRI based radiomics will be used to solve the preoperative and non-invasive diagnosis and prognosis prediction of glioma and liver cancer.Firstly,the MRI-based radiomics analysis was performed to non-invasively predicting chromosome 1p/19q co-deletion in World Health Organization(WHO)grade II and III(lower-grade)gliomas before surgery.This retrospective study included 277 patients histopathologically diagnosed with lower-grade glioma,and a total of 647 radiomic features were extracted from each patient's imaging.A set of 64 radiomic features was selected using Variance Threshold and univariate analysis.Then,we applied the random forest algorithm to generate a radiomics signature for predicting 1p/19q co-deletion in the training cohort.In addition,the clinical model and combination model was also constructed using a logistic regression algorithm.The radiomics signature outperformed the clinical model and combination model,with areas under curve(AUCs)of 0.887 and 0.760 in raining and validation cohorts,respectively.The Delong test revealed significant differences between the radiomics signature and the clinical model(p=0.005 in validation cohort).Our study highlighted that radiomics features can be used to pre-operatively and non-invasively distinguish the 1p/19q co-deletion genotype in patients with lower-grade gliomas.Secondly,we conducted a multi-habitat radiomics study based on multi-parameter MRI to effectively and accurately predict the histopathologic growth patterns(HGPs)of colorectal liver metastases(CRLMs),so as to effectively screen patients and develop appropriate treatment strategies.This study included 182 resected and histopathological proven CRLMs of chemotherapy-naive patients from two institutions.A total of 828 radiomic features were extracted from two sets of regions of interest(ROI),the tumor zone and the tumor–liver interface(TLI)zone,respectively.Radiomics analysis was performed separately in two sets of radiomic features.The robust feature selection(RFS)method was used to rank the coefficients of all features and top 20 best features were reserved as the effective predictors for the later decision tree classifier in each sequence.Then,the forward stepwise was used to select the desired sequences from five sequence,and the final radiomics signature was generated based on the desired sequences through logistics regression method.The radiomics^TLImodel exhibited better performance than radiomics^tumor model,with AUCs of0.974,0.912 and 0.974 in training,internal validation cohort and external validation cohort,respectively.In addition,the clinical and combination models were developed through multivariate logistic regression method.The combination model exhibited better performance than that of clinical model,which indicated that the radiomic features can not only be used to identify the HGPs of CRLMs,but also complement the clinical factors in predicting HGPs of CRLMs.Finally,a CT-based radiomics model for preoperative prediction of recurrence-free survival(RFS)of patients with intrahepatic cholangiocarcinoma(ICC)after radical resection were developed.In order to fully mine the heterogeneity information,two sets of imaging features:512 deep learning features that based residual network(Res Net)and 473 artificially designed radiomics features were extracted.A deep learning signature and a radiomics signature were generated based on the deep learning features and radiomic features selected by least absolute shrinkage and selection operator(Lasso)-Cox algorithm,respectively.The deep learning model exhibited better performance than that of radiomics model with C-indexes of validation group were 0.656 and 0.615,respectively.In addition,a clinical model and three combination models were constructed through Cox proportional hazard regression algorithm.The combination model that integrated clinical factors,deep learning signature and radiomics signature obtained best predictive performance(C-index of 0.809 and 0.819 in training and validation cohort,respectively).It suggested that clinical factors and imaging features are complementary in predicting RFS of patients with intrahepatic cholangiocarcinoma.The nomogram that based on combination model were constructed to predict the recurrence probability in 3-month,6-month and 1-year,so as to optimize the treatment strategy of patients.In addition,all patients were successfully classified as high or low risk groups according to the risk score generated through combination model,thus benefit patient management.