Construction And Validation Of A Radiation Pneumonia Prediction Model Based On Radiomics And Multiple Parameters

Objective(s):In this study,a multi-parameter prediction model was constructed by combining clinical and dosimetric parameters with the extracted imaging features from the training set of computer tomography(CT)scans,and the joint model was visualized.The potential of the model to predict the probability of radioactive pneumonia(RP)after standard lung radiotherapy in patients with non-small cell lung cancer(NSCLC)was evaluated by the validation set.Methods:Patients with a pathologically confirmed diagnosis of non-small cell lung cancer and considered non-viable for surgery after a multidisciplinary consultation were collected from January 2008 to December 2017 at our institution,and all patients received complete and standard pulmonary radiotherapy.Clinical data,laboratory tests,localization scans of the chest and radiotherapy dosimetry were retrospectively collected from patients.The region of interest(ROI)was extracted from the lung tumor and the tumor peri-tumor extension of 0.5 cm on the pre-radiotherapy localization CT images,and the intratumor prediction model and peri-tumor prediction model were constructed,and then the most meaningful prediction model was evaluated and compared,and the imaging histology score of the model was calculated.At the same time,single-factor and multi-factor logistic regression methods were used to screen out clinically relevant predictors and dosimetric parameters to jointly construct a joint model,and finally the joint model was validated for its predictive efficacy and clinical benefits and visualized.Results:A total of 104 patients with inoperable NSCLC were included in the study,including 59 patients with radiation pneumonia within 6 months after radiotherapy and 45 patients without radiation pneumonia.All patients were segmented in equal proportions by stratification,with 73 cases in the training set and31 cases in the test set.Comparing the area under the receiver operating characteristic curve(AUC)of the intratumoral model with that of the peri-tumoral model,the intratumoral model had an AUC of 0.871,a specificity of 0.771,and a sensitivity of0.868 in its training set,and an AUC of 0.719,a specificity of 0.400,and a sensitivity of0.952 in its test set.In the test set,the AUC value was 0.719,the specificity was0.400,and the sensitivity was 0.952.The peritumoral model had an AUC of 0.798,a specificity of 0.629,and a sensitivity of 0.921 in its training set,and an AUC value of0.714,a specificity of 0.500,and a sensitivity of 0.857 in the test set.Comparison of the AUC of the two models revealed that the predictive efficacy of the intratumoral model was better than that of the peritumoral model.In addition,among the clinical parameters,patient age,smoking history,and lymphocyte ratio were considered to be statistically different by multifactorial analysis.Dosimetric parameters of mean lung dose(MLD)and percent volume receiving at least 20Gy(V20)were analyzed by univariate analysis.Therefore,the joint model uses the radiomic scores established by the intratumoral model in conjunction with the above-mentioned multiple parameters to build a prediction model,the ROC of this a combined prediction model model has an AUC of 0.928,a specificity of 0.914,and a sensitivity of 0.789 in the training set,and an AUC of 0.629,a specificity of 0.70,and a sensitivity of 0.619 in the test set.and the clinical benefit of the model is validated by calibration curves and decision curve analysis.Conclusion(s): A combined prediction model based on CT images outlining tumor ROI screening radiomic features combined with multi-parameter construction can predict the probability of radiation pneumonia after radiotherapy in inoperable NSCLC patients,and the model has some clinical value to provide a basis for early clinical intervention in RP.