Analysis Based On Multimodality Radiomics Prediction Model Of Distant Metastasis Of Locoregionally Advanced Nasopharyngeal Carcinoma

Purpose: Nasopharyngeal carcinoma(NPC)is aggressive cancer with various biological behaviors.At present,treatment of nasopharyngeal carcinoma is mainly determined by TNM staging system.Diverse response to radiation is not reflected in TNM system.Patients with the same stage have different biological behaviors and different prognosis and outcomes after the same treatment.This study provides more evidences to guide personalized treatment for those patients.Method: A total of 764 nasopharyngeal cancer patients admitted to Hunan Cancer Hospital from January 2013 to December 2017 were selected retrospectively.The patients were followed up for the location of distant metastasis and the time when the distant metastasis was found.Clinical data were collected,including age,gender,pathological type,T stage,N stage.DICOM images of CT,T1 WI,T2WI,and enhanced T1 WI MRI were received in Hunan Cancer Hospital.After the image is preprocessed,IBEX is used to extract the radiomics features.The random forest method dimensionally reduces the features.A five-fold cross-validation method was used to divide the training set and the test set.Use logistic regression,random forest,support vector machine and extra random trees to train and verify the single-modality and multi-modality prediction model.Single-modality model includes clinical model,CT-based image model and MRI-based image model;multi-modality model includes clinical combined CT-based image model,clinical combined MRI-based model,and clinical combine CT and MRI-based model.The sensitivity and specificity of each model were calculated respectively,and the AUC value of receiver operating curve was used to evaluate the effectiveness of the model to verify the model.Result: A total of 90 patients had distant metastases within 5 years after treatment.In this study,6 distant metastasis prediction models were trained and validated using clinical data,CT,and MRI imaging features of 764 patients.1646 radiomics features were extracted from images of different modalities respectively,10 features of each model related to distant metastasis were retained after feature selection,namely the T stage in Clinical Feature,maximum 3D diameter in Shape,global standard error in Intensity Histogram,and entropy,contrast,cluster shade,correlation,inverse difference matrix,and inverse variance in GLCM.The prediction model of clinical data combined with multimodal images has the best performance(accuracy of 0.73±0.01,AUC of 0.75±0.02).Conclusion: In this study,a multi-modality radiomics model was established for predicting distant metastasis after nasopharyngeal carcinoma treatment and proved that its efficacy is superior to both traditional clinical features and single-modal imaging omics models.These features can be used as image biomarkers to indicate the possible outcomes of patients with nasopharyngeal cancer,and provide doctors and patients with more information in the future to optimize the personalized treatment of nasopharyngeal cancer.