| Background and purpose: For patients with the early-stage non-small cell lung cancer(NSCLC),whose tumors behavior the high malignancy and large heterogeneity,postoperative recurrence and metastasis will result in treatment failure and shorted survival even after the standardized surgical resection.It emphasizes the need for a biomarker that could identify patients with high-risk recurrence for more aggressive follow-up and individualized treatment making.The pathological TNM staging is considered as the most important prognostic factor for NSCLC,but there remains the various survival for those in same stage.Recently,CT-based radiomics,as a comprehensive technology of image information mining,can quantify the intensity,texture,and shape-related information from tumor,which could be used for developing a biomarker for survival risk assessment.As such,this study aimed to construct a radiomics-based prognostic biomarker for early-stage NSCLC,explore its effectiveness and clinical usefulness beyond those of the traditional clinicopathologic risk factors,investigate the performance difference between the biomarker derived from single-and multi-dimensional radiomics features,and explore the genetic interpretability of the radiomics-based prognostic marker.Materials and Methods: This study retrospectively retrieved the patients who underwent surgical resection due to clinical stage I NSCLC between January 2011 and December 2013 at our institution,and their clinicopathologic data and CT images were collected.With the same inclusion and exclusion criteria,patients in the TCIA Radiogenomics were included as an external validation set.For CT image processing,we firstly visually collected 11 imaging semantic features,such as imaging subtype,lobulation,spiculation,pleural invasion,and etc.Secondly,an open-source platform,3D-slicer software,were used to manually segment the intratumoral 3D region of interest(ROI)from CT images,and then 107 original radiomics features were extracted with Pyradiomics package.To investigate the predictive value of radiomics derived from different region,an intratumoral 2D ROI and the peritumoral ROI were automatically generated based on the intratumoral 3D ROI,in which a total of 2163 radiomics features were extracted from the three ROIs after adding a wavelet transform filter to the feature extraction process.After feature selection with intraclass correlation coefficient analysis,univariate survival analysis and LASSO COX regression model,the most significant radiomics features and their weight coefficients were fitted to construct prognostic signatures.The recurrence-free survival(RFS)and overall survival(OS)among different risk groups defined by clinicopathological factors,semantic features or imaging markers were compared using the Kaplan-Meier method and log-rank test.Univariate and multivariate Cox regression analysis were used to identify independent predictors of RFS and OS.The discriminative abilities of these prognostic factors were quantified using the area under the time-dependent receiver operating characteristic curve(AUC)or the Harrell concordance index(C-index),which were compared with De Long tests.A weighted gene co-expression network was constructed based on the RNA sequencing data of 35 patients in the external validation set.Gene module was clustered and their association with the radiomics features was explored.We used gene set variation analysis to explore the functional pathways related to imaging features in gene modules,and Hub genes were further screened from co-expression network by TOM.Finally,the underlying mechanism of Hub gene affecting the prognosis of early-stage NSCLC was explored based on cell functional experiments.Results: The main results are as follows:1)The identification of traditional clinical-imaging prognostic factors and their efficacy in early-stage NSCLC: In this study,1461 clinical stage I NSCLC patients were included,comprising 338 subsolid lesions and 1123 pure-solid lesions.The recurrence(35.4%)and death(28.1%)rate in pure-solid NSCLC were significantly higher than those in sub-solid NSCLC(p<0.001),for whom the 5-year RFS was95.2% and 66.4%,and the 5-year OS was 96.6.% and 74.6%,respectively.When exploring the interactive prognostic effect of imaging type and T stage,we found that there was a moderate difference in RFS of subsolid NSCLC patients among T stages(p=0.045),and no significant difference in OS(p=0.260)absolutely.For those with pure-solid NSCLC,there was a significant survival difference among different T stages(all p<0.001).Multivariate analysis confirmed that the imaging subtype remained an independent and significant predictor of RFS(p=0.001)and OS(p<0.001)after adjusted by age(p=0.001 and p<0.001)and pathological lymph node invasion(p=0.003 and 0.044).In addition,the lobulation(p=0.046)was also an independent risk factor of death.The above two independent clinicopathological features,age and pathological lymph node invasion,predicted the 5-year RFS and OS with AUCs of0.661 and 0.650,respectively;and the AUCs of imaging features were 0.614 and0.597,respectively;Integrating the above clinicopathological and imaging features reached the highest AUCs with 0.709 and 0.698,respectively.2)Multiregional and multidimensional radiomics features improved the prognostic value of traditional clinicopathological factors in pure-solid early-stage NSCLC: Radiomics analysis was performed in 592 multi-cohort patients with stage IA pure-solid NSCLC,which was divided into the training set(TS,n=381),internal validation set(IVS,n=163),and external validation set(EVS,n=48).Based on the features extracted from multi-regional and multi-dimensional ROIs,8 intratumoral3D-ROI features,5 intratumoral 2D-ROI features and 5 peritumoral features were screened out and fitted into a signature for survival risk stratification,which yielded the highest 3-year and 5-year AUCs of 0.77 and 0.78 in the IVS and 0.76 and 0.75 in the EVS,respectively.The 5-year RFS rates of patients in the low-and high-risk groups were 93.4% and 70.5% in the IVS,95.2% and 47.3% in the EVS.Meanwhile,there was no statistically significant difference of RFS between the low-risk pure-solid group defined by the radiomics signature and the subsolid NSCLC(p>0.05).Multivariable analysis suggested that the multi-scale radiomics signature remained an independent prognostic factor(hazard ratio: 6.2;p<0.001),and improved the discriminative ability(C-index: 0.75 vs.0.59)and clinical usefulness of conventional clinicopathological factors(net reclassification index: 31%;integrated discriminant improvement: 13.2%).3)Genetic-imaging analysis and cell function experiment revealed the interpretation of the multiregional radiomics signature: 31 gene modules were clustered and identified by the Weighted gene correlation network analysis(WGCNA)according to the gene expression pattern.The correlation assessment between the functional modules and radiomics features suggested that the Saddlebrown module was highly correlated with 4 radiomic features,while the Lightcyan module was positively correlated to 3 features,and the Green module was negatively correlated to the radiomics signature.Gene set variation analysis(GSVA)indicated that some specific features were associated with cell proliferation,apoptosis,migration and invasion,cellular energy metabolism and endoplasmic reticulum stress regulation.The co-expression network constructed based on the TOM matrix further screened out9 hub genes that were significantly associated with the radiomics features,such as CDC20,DAPK2,DAPK2 and etc.The cell functional experiments suggested that the low expression of CDC20 significantly inhibited the proliferation of lung cancer cells,the low expression of DAPK2 promoted the survival of lung cancer cells and inhibited their apoptosis,and the silencing of F2RL1 highly inhibited the migration and invasion of lung cancer cells.Conclusion: The semantic features,such as image subtype and lobulation,were independent predictors of postoperative recurrence or death beyond the clinicopathological feature,but remained a relative low efficiency.For patients with pure-solid NSCLC,the radiomics signature based on multiregional features yielded the best discriminative ability of prognosis assessment,which could reflect the underlying biological process of lung cancer cell proliferation,apoptosis,migration and invasion. |
PDF Full Text Request