Prediction Of Response To Neoadjuvant Chemotherapy In Cervical Cancer Using Multicenter CT And MRI-based Radiomic Analysis

BackgroundPreoperative neoadjuvant chemotherapy(NACT)is widely applied as major treatment in routine practice for Chinese locally advanced cervical cancer(LACC)patients due to its ability to reduce the tumour volume and render unresectable tumors operable.However,chemotherapeutic non-responders experience unnecessary chemotherapy-related toxicities and disease progression due to delay in effective treatment,which worsen the prognosis.This scenario highlights the need for accurate identification of chemotherapeutic non-responders and responders prior to the initiation of therapy.However,no effective and clinically validated predictor of chemotherapeutic responsiveness has been established.Radiomics,an emerging approach for image interpretation,can non-invasively predict treatment response solely by radiographic examination,showing promising for pre-therapy prediction of response to NACT for LACC.Thus,in this project we aimed to investigate whether radiomic features based on pre-therapeutic imaging data could predict the clinical response to neoadjuvant chemotherapy(NACT)in LACC patients.Methods and Results(1)Prediction of response to preoperative neoadjuvant chemotherapy in locally advanced cervical cancer using multicenter CT-based radiomic analysis277 LACC patients treated with NACT followed by surgery/radiotherapy from 10 different hospitals were included in this retrospective study.One thousand and ninety-four radiomic features were extracted from venous contrast enhanced and non-enhanced CT imaging for each patient.Five combined methods of feature selection including support vector machine,lasso(LASSO),ridge regression,random forest,extreme random trees were used to reduce dimension of features.6-features radiomics signature was constructed by Random Forest(RF)method while a combined model incorporating radiomics signature with clinical factors was developed by multivariable logistic regression.Radiomics signature containing pre-and post-contrast imaging features could adequately distinguish chemotherapeutic responders from non-responders in both primary and validation cohorts[AUC:0.773(95%CI,0.701-0.845)and 0.816(95%Cl,0.690-0.942),respectively].The combined model had a better predictive performance with an AUC of 0.803(95%CI,0.734-0.872)in the primary set and an AUC of 0.821(95%CI,0.697-0.946)in the validation set,compared to radiomics signature alone.Besides,the combined model showed good calibration and clinical usefulness.(2)MRI-based radiomic analysis for pretreatment prediction of response to neoadjuvant chemotherapy in locally advanced cervical cancer:a multicentre studyA total of 275 LACC patients from 8 different hospitals receiving NACT were enrolled in this study,which were allocated to the training and testing sets(2:1 ratio).Three radiomic feature sets were extracted from intratumoral region of T1-weighted images,intratumoral and peritumoral region of T2-weighted images before NACT for each patient.All features were selected by support vector machine recursive elimination(SVM-RFE)method.Then three single sequence radiomic models and three combined models which integrated features of different regions or sequences were constructed.The performance of the six models were assessed using receiver operating characteristic curve.The three sequence combined model of the intratumoral region of T1-weighted images,intratumoral and peritumoral regions of T2-weighted images had the best predictive performance achieving an AUC of 0.998 in the training set and 0.999 in the testing set,which was significantly better than the other models.Meanwhile,the three sequence model also had the best predictive accuracy,sensitivity and specificity.(3)External validation of CT and MRI-based radiomics models for pretreatment prediction of response to neoadjuvant chemotherapy in LACC patientsIn order to evaluate the robustness and consistency of the established CT and MRI-based radiomics model,additional external validation groups consisted of different training sets by randomly combinations of different hospitals.Through the same feature selection and model construction process as previous experiment,the robustness of the constructed models for NACT response prediction was evaluated by the ROC curve.For CT-based radiomics signature,Only<4%difference was observed between the two validation groups and both groups showed good classification performance,consistent to the original model(0.796,0.822).For MRI-based radiomics combined model,ROC curves of the three various external validation groups were also consistent with no significantly difference.Similarly,three-sequence model still had the highest sensitivity and specificity,outperformed the other two-sequence models.This result revealed that both CT and MRI-based radiomics model exhibited robust and persistent good predictive performance between different centers regardless of patient composition of the training set.ConclusionNewly developed CT or MRI-based radiomic model had good predictive performance for NACT response and remained relatively stable across centers.This easy-to-use predictor of chemotherapeutic response with improved predictive ability,might facilitate treatment strategies tailored for individual cervical cancer patients.