Prediction Of Progression-free Survival And Pathological Types Of Locally Advanced Cervical Cancer After Concurrent Chemoradiotherapy Based On PET-CT Radiomics

Objective:1)To explore the prognostic value of baseline 18F-fluorodeoxyglucose（FDG）PET/CT metabolic parameters in locally advanced cervical cancer（LACC）after concurrent chemoradiotherapy（CCRT）.2)To construct machine learning models based on CT imaging and clinical parameters for predicting progression-free survival（PFS）of LACC patients after CCRT.3)The aim of this study was to develop and validate radiomics models based on pretreatment ¹⁸F-FDG PET/CT images to estimate the probability of PFS in LACC.4)To distinguish cervical adenocarcinoma（AC）from cervical squamous cell carcinoma（SCC）using¹⁸F-FDG PET/CT-based radiomics features.Methods:1)From September 2015 to October 2021,the clinical data of 180 LACC patients（age:22-76 years）who underwent¹⁸F-FDG PET/CT before CCRT at Affiliated Cancer Hospital of Shandong First Medical University were analyzed retrospectively.The metabolic tumor volume（MTV）,total lesion glycolysis（TLG）,maximum standardized uptake value(SUV_max),and mean standardized uptake value(SUV_mean)were computed by using the margin threshold of 42%SUV_max.The optimal threshold for predicting PFS was obtained by ROC curve analysis,the Kaplan-Meier was applied for survival analysis,and the log-rank test was applied to compare the survival rate between groups.Multivariate Cox proportional hazard regression was used to analyze predictors for PFS.2)Clinical data of 167 LACC patients treated with CCRT at Shandong Cancer Hospital from September 2015 to October 2021 were retrospectively analyzed.All patients were randomly divided into the training and validation cohorts according to the ratio of 7 vs.3.Clinical features were selected by univariate and multivariate Cox proportional hazards model.Radiomics models and nomogram were constructed by radiomics features which were selected by least absolute shrinkage and selection operator（LASSO）Cox regression model to predict the 1-,3-and 5-year PFS.Combined models and nomogram models were developed by selected clinical and radiomics features.The Kaplan-Meier curve,receiver operating characteristic（ROC）curve,C-index and calibration curve were used to evaluate the model performance.3)A total of 190 LACC patients received concurrent radiochemotherapy were enrolled in this study from two hospitals,and randomly allocated to training cohort（n=117）and validation cohort（n=50）according to the ratio of 7:3 from one hospital and another as the external validation cohort（n=23）.Radiomics models were constructed by the LASSO Cox model based on pretreatment¹⁸F-FDG PET/CT images.Clinical features were evaluated by univariate and multivariate Cox proportional hazards model.Radiomics-clinical model was developed by selected clinical and radiomics features.The predictive power was assessed by the receiver operating characteristic curve,Kaplan-Meier curve and independent dataset.4)Pretreatment ¹⁸F-FDG PET/CT were retrospectively collected from patients who were diagnosed with SCC or AC at Shandong Cancer Hospital from September 2015 to May 2022.All patients underwent ¹⁸F-FDG PET/CT imaging before treatment.Radiomics features were extracted based on baseline¹⁸F-FDG PET/CT images,and Spearman correlation coefficient and LASSO were used to select radiomics features.Six machine learning algorithms were then applied to establish models,and the best-performing classifier was selected based on accuracy,sensitivity,specificity,and area under the curve（AUC）.Results:1)The median follow-up was 19.1months,and 54 patients（30.0%,54/180）suffered from disease progression.ROC analysis showed that the optimal cut-off value of MTV was 31.145 ml,with the AUC of 0.641.Para-aortic lymph node（PALN）metastasis had the highest AUC value（0.589）among the clinical factors,followed by international federation of gynecology and obstetrics（FIGO）stage（0.581）.The 1-year PFS rates of patients with MTV<31.145ml（n=88）and MTV≥31.145ml（n=92）were 80.68%and 59.78%,respectively（X²=13.72,P<0.001）.Multivariate Cox analysis demonstrated that pathological type（hazard ratio（HR）=3.075,95%CI:1.370-6.901,P=0.006）,FIGO stage（HR=1.955,95%CI:1.031-3.707,P=0.040）,PALN metastasis（HR=2.136,95%CI:1.202-3.796,P=0.010）and MTV（HR=2.449,95%CI:1.341-4.471,P=0.004）were the significant predictors for PFS.2)A total of 1 409radiomics features were extracted based on the region of interest（ROI）in CT images.CT radiomics models showed better performance for predicting 1-,3-and 5-year PFS than the clinical model in the training and validation cohorts.The combined model displayed the optimal performance in predicting 1-,3-and 5-year PFS in the training cohort（AUC:0.760,0.648,0.661,C-index:0.740,0.667,0.709）and validation cohort（AUC:0.763,0.677,0.648,C-index:0.748,0.668,0.678）.3)Three CT radiomics features,one PET radiomics feature and one clinical feature were significantly associated with PFS in the training cohort.The radiomics-clinical model displayed better performance in training,validation and external validation cohorts for prediction 3-year PFS（AUC,0.661,0.718,0.765.C-index,0.698,0.724,0.693）and 5-year PFS（AUC,0.661,0.711,0.719.C-index,0.698,0.722,0.766）.Decision curve analysis confirmed their clinical usefulness.4)The PET radiomics model constructed based on the lightGBM algorithm had an accuracy of 0.915 and an AUC of 0.851（95%CI,0.715-0.986）in the validation cohort,which were higher than that of the CT radiomics model（accuracy:0.661;AUC:0.513[95%CI,0.339-0.688]）.The DeLong test revealed no significant difference in AUC between the combined radiomics model and the PET radiomics model in both the training cohort（P=0.347）and the validation cohort（P=0.776）.Conclusion:1)MTV,Pathological type,FIGO stage and PALN metastasis are independent prognostic risk factor for PFS,and MTV as the baseline ¹⁸F-FDG PET/CT metabolic parameter is superior to clinical factors in predicting PFS.2)Combined model constructed based on CT radiomics and clinical features yield better prediction performance than that based on radiomics or clinical features alone.As an objective image analysis approach,it possesses high prediction efficiency for PFS of LACC patients after CCRT,which can provide reference for clinical decision-making.3)Different CT radiomics features have a certain timeliness difference in cervical cancer survival prediction.The combined model and nomogram based on¹⁸F-FDG PET/CT images and pathology may be a clinically applicable tool for the early assessment of the long-term prognosis of patients diagnosed with LACC treated with definitive chemoradiotherapy.4)The Light GBM-based PET radiomics model had great potential to predict the fine histological subtypes of LACC,which might serve as a promising noninvasive approach for the diagnosis and management of locally advanced cervical cancer.