| BackgroundLung cancer is the leading cause of cancer-related mortality worldwide and non-small cell lung cancer(NSCLC)accounts for about 85%of all lung cancers.Surgical resection with a curative intent is regarded as the cornerstone of treatment for stage ? to stage ?a NSCLC,while the prognosis can vary greatly among patients operated for the tumor resection.Clinically,patients need to be accurately assessed about the prognosis after surgery and stratified as high-or low risk of recurrence or death,which guides personalized treatment and follow-up programs and assists clinical decision-making.Currently,prognostic evaluation of NSCLC is mainly based on the TNM stage established by American Joint Committee on Cancer(AJCC)(T,tumor,N,lymph node,M,distant metastasis).However,the wide spectrum of survival time exists in the same-staged NSCLC even after complete resection,which demonstrates the urgent need for personalized medicine.Recently,radiomics,extracting high throughput quantitative descriptors from routinely acquired computed tomography(CT),enables the noninvasive profiling of tumor heterogeneity and assists clinical decision-making.Radiomics provides new ideas and directions for the prognosis estimation of patients with NSCLC.ObjectiveIn order to assist doctors in arranging the postoperative treatments and re-examinations for NSCLC patients,this study was initiated to develop a hand-crafted and deep learning feature signature based on CT images to estimate overall survival(OS)in patients with NSCLC,respectively.A nomogram incorporated hand-crafted feature signature,deep learning feature signature with clinicopathologic risk factors was constructed to assess the value for individual OS estimationMaterials and MethodsThe data of 412 NSCLC patients(275 patients for training dataset,137 for validation dataset)were retrospectively collected to predict the patients' prognosis survival.Firstly,lung tumors were manually segmented on the largest cross-sectional area of tumor by using Image J software,and the hand-crafted features and deep learning features were extracted using Matlab-based feature extraction software,respectively.Secondly,the hand-crafted and deep learning features associated with OS were screened by using the least absolute shrinkage and selection operator(LASSO)Cox regression model analysis,respectively.Finally,a hand-crafted feature signature and a deep learning feature signature were generated by using the selected features to predict the prognosis survival of NSCLC patients,respectively.Further validation of the signatures as an independent biomarker was performed by using multivariate Cox regression method.A combined nomogram with the hand-crafted signature,deep learning feature signature and clinicopathologic factors was constructed to estimate the individual OS,which was then assessed with respect to discrimination,calibration and clinical usefulness.ResultsThe 21-hand-crafted feature signature and 19-deep learning feature signature were significantly associated with OS,respectively(?0.013),independent of clinicopathologic risk factors(HR,2.006[95%CI,1.115-3.610]for training dataset,2.187[95%CI,1.392-3.430]for validation dataset).Incorporating both signtures into the combined nomogram resulted in good discriminiation for the estimation of OS(C-index=0.785 for training dataset,0.800 for validation dataset),as well as a good calibration.Decision curve analysis demonstrated that in terms of clinical usefulness,the combined nomogram had a good overall net benefit across the majority range of reasonable threshold probabilities.ConclusionCT radiomics features including hand-crafted and deep learning features could effectively assist doctors to make more accurate in prognostic survival prediction for NSCLC patients,and the combination of the features with conventional clinicopathological features can individually estimate prognosis,so as to help doctors to optimize treatment and follow-up plans for NSCLC patients to extend their survival time and improve the survival rate. |
PDF Full Text Request