MR-based Radiomics For Precise Prediction Of Lymph Node Metastasis And Recurrence Risk Stratification In Early-stage Cervical Squamous Cell Carcinoma

BackgroundSurgery is the preferred treatment for early-stage cervical cancer,and adjuvant treatment is indicated if risk factors of recurrence are discovered after operation.Studies have shown that,surgery and radiotherapy have similar rates of effectiveness,and multimodality therapy can increase toxicity.Therefore,it is recommended that radiotherapy is the first choice for patients who may need adjuvant therapy after surgery.However,there remains an unmet clinical need for the precise indicators to quantify the recurrence risk stratification for patients with early-stage cervical cancer.In recent years,it has been confirmed that radiomics has the ability to capture intratumour heterogeneity in a non-invasive way,which provides a new idea for preoperative precise prediction of risk factors and risk stratification in patients with early-stage cervical cancer.PurposeTo determine the value of a radiomics model incorporating the handcrafted signature and deep learning signature for preoperative prediction of lymph node metastasis in early-stage cervical squamous cell carcinoma(CSCC),and to explore the clinical value of MR-based radiomics for recurrence risk stratification in early-stage CSCC.Materials and methodsPart 1:A total of 190 eligible patients with early-stage CSCC were randomly divided into training(n=100)and test(n=90)sets.Handcrafted features and deep learning features were extracted from T2-weighted fat suppression sequence.Interclass correlation coefficients,C-statistics,the minimum redundancy maximum relevance algorithm(mRMR)and LASSO regression with 10-fold cross-validation were used for key features selection,and the key features were used to constructed handcrafted signature and deep learning signature by multivariate logistic regression.A combined model that incorporated the handcrafted signature,deep learning signature and squamous cell carcinoma antigen(SCC-Ag)levels was developed by logistic regression.The model performance was assessed and validated by calibration curve,ROC and decision curve,with respect to its calibration,discrimination and clinical usefulness.Part 2:A total of 218 eligible patients with early-stage CSCC were divided into training(n=145)and test(n=73)sets.Recurrence risk was stratified into low-risk status and positive-risk status,according to the postoperative pathological results of 3 high-risk factors and 3 intermediate-risk factors(Sedlis criteria).Radiomics features were extracted from T2-weighted fat suppression and T1 contrast-enhanced sequences,respectively.The key features were selected and then used to constructed T2-rad(radiomics signature based on T2-FS squence)and CS-rad(radiomics signature based on sagittal T1+C sequence)by multivariate logistic regression.A combined model that incorporated the T2-rad,CS-rad and SCC-Ag was developed by logistic regression.The model performance was assessed and validated by calibration curve,ROC and clinical impact curve,with respect to its calibration,discrimination and clinical usefulness.ResultsPart 1:Three handcrafted-features and 3 deep learning-features were selected and used to build handcrafted signature(HC-rad)and deep learning signature(DL-rad).The combined model,which incorporated the HC-rad,DL-rad and SCC-Ag showed satisfactory discrimination and calibration in the training set(AUC=0.852,95%CI=0.761-0.943)and the validation set(AUC=0.815,95%CI=0.711-0.919).The combined model yielded greater AUCs than either the HC-rad(AUC=0.794,0.725,respectively),DL-rad(AUC=0.735,0.718,respectively)or the SCC-Ag(AUC=0.735,0.688,respectively)alone in both the training and test sets.Decision curve analysis indicated the clinical usefulness of the combined model.Part 2:Four key radiomics features were selected from each of the two sequences.Multivariate analysis showed that SCC-Ag was not an independent risk predictor(p=0.402),and both T2-rad and CS-rad signatures were independent risk predictors(p<0.05).A prediction model integrating the T2-rad and CS-rad signature was developed.The combined model showed satisfactory discrimination and calibration in the training set(AUC=0.832,95%CI=0.742-0.932)and the test set(AUC=0.837,95%CI=0.763-0.902).The combined model yielded greater AUCs than either the T2-rad(AUC=0.810 and 0.796,respectively)or the CS-rad(AUC=0.793 and 0.781,respectively)in both the training and test sets.Decision curve analysis indicated the clinical usefulness of the combined model.ConclusionA combined model integrating handcrafted signature and deep-learning signature demonstrated good performance in identifying lymph node metastasis in patients with early-stage CSCC.And MR-based radiomics could be used for risk stratification in patients with early-stage CSCC,which was expected to be an auxiliary tool for clinical treatment decision-making.