Radiomics For Prediction Of ALK Mutations,Ki67 Expression And Prognostication In Non-Small Cell Lung Cancer

Objectives1.To investigate the value of preoperative CT single-task radiomics signature for individually predicting ALK mutation status,Ki-67 expression in non-small cell lung cancer(NSCLC).2.Developed and verified the model based on preoperative CT multi-task radiomics signature for simultaneously predicting ALK mutation status and Ki-67 expression level in NSCLC.3.Developed and verified a nomogram based on the preoperative CT deep learning signature in prognostic prediction of NSCLC.Methods1.370 patients in the training group and 135 patients in the validation group.CT radiomics features extraction was performed by SERA and the least absolute contraction.The selection operator(LASSO)was used to screen out features and then constructing the radiomics signature.The radiomics signature wea combined with independent clinical risk factors to construct a predictive model.Model performances were evaluated through discrimination,calibration,and clinical usefulness.Net Reclassification index(NRI)was calculated to quantify the incremental value of radiomics features.The results were verified in the validation cohorts.2.Research objects and feature extraction methods were same as above.We use multitasking elastic network regression to select all the tasks related public feature subset.The multitask radiomic features were used to construct radiomics signature.Univariate and multivariate regression analyses were used to select clinical indicators associated with ALK mutation status and Ki67 expression level,then combing the above clinical indicators with radiomics signature to establish prediction model.The model evaluation and validation method were the same as above.3.Deep learning features were extracted from non-enhanced and venous-phase CT images for each NSCLC patient in training cohort(n=231).A deep learning signature was built with LASSO Cox regression model.Combing the signature and other independent clinical risk factors to construct the nomogram.The model evaluation and validation method were the same as above.Results1.ALK mutation status of NSCLC was predicted by a model combining single-task radiomics signature and age(AUC training=0.75;AUC validation=0.76).The radiomics signature had incremental value(NRI training=0.27,P=0.01;NRI validation=0.10,P=0.46).2.Ki67 expression of NSCLC was predicted by a model combining single-task radiomics signature,gender and pathological type(AUC training=0.75;AUC validation=0.75).The radiomics signature had incremental value(NRItraining=0.20,P<0.001;NRI validation=0.08,P<0.001).3.ALK mutation status(AUC training=0.78;AUC validation=0.81)and Ki67 expression level(AUC training=0.80;AUC validation=0.79)were predicted by the model combining age,gender,pathological type and multi-task radiomics signature were better than single-task predicting model and clinical model4.Combining deep learning signature with clinical risk factors of TNM stage,lymphatic vessel invasion and differentiation grade showed reasonable discriminative ability(C-index=0.80)and reflected a good calibration,which was validated in external validation cohort(C-index=0.72).Additional value of deep learning signature to the nomogram was statistically significant(NRI=0.09,P=0.03 for training cohort;NRI=0.11,P=0.04 for validation cohort).Conclusions1.ALK mutation status and Ki67 expression in NSCLC can be predicted individually by the models based on the preoperative CT single-task radiomics signatures.2.Joint prediction of Ki-67 expression and ALK mutation status in NSCLC based on multi-task radiomics signature perform better than single-task predicting model and clinical model.3.Deep learning has a good application value in NSCLC prognostic prediction,it provides a practical reference for clinical treatment decision making.