Prediction of epidermal growth factor receptor mutation subtypes in patients with non-small cell lung cancer by 18F-FDG PET/CT radiomics / 中华核医学与分子影像杂志

ABSTRACT

Objective:To investigate the value of pre-therapy 18F-FDG PET/CT radiomic models in differentiating epidermal growth factor receptor (EGFR) exon 19 deletion from exon 21 L858R missense in patients with non-small cell lung cancer (NSCLC). Methods:A total of 172 patients with EGFR mutant NSCLC (54 males, 118 females, age: (56.2±12.5) years) in the Fourth Hospital of Hebei Medical University between January 2015 and November 2019 were retrospectively included. Exon 19 mutation was found in 75 patients and exon 21 mutation was identified in 97 patients. The patients were divided into training set ( n=121) and validation set ( n=51) in a 7∶3 ratio by using random number table. The LIFEx 4.00 package was used to extract texture features of PET/CT images of lesions. The least absolute shrinkage and selection operator (LASSO) algorithm was used for feature screening. Three machine learning models, namely logistic regression (LR), random forest (RF), and support vector machine (SVM) models, were constructed based on the selected optimal feature subsets. The ROC curve analysis was performed to assess the predictive performance of those models. Finally, decision curve analysis (DCA) was used to evaluate the clinical value of the models. Results:Nine radiomics features, including 6 PET features (histogram (HISTO)_Kurtosis, SHAPE_Sphericity, gray level run length matrix (GLRLM)_ low gray-level run emphasis (LGRE), GLRLM_ run length non-uniformity (RLNU), neighborhood grey level different matrix (NGLDM)_Contrast, gray level zone length matrix (GLZLM)_ short-zone low gray-level emphasis (SZLGE)), and 3 CT features (gray level co-occurrence matrix (GLCM)_Correlation, GLRLM_ run percentage (RP), NGLDM_Contrast), were screened by LASSO algorithm. Three machine learning models had similar predictive performance in the training and validation sets: AUCs for the RF model were 0.79, 0.77, and those for the SVM model were 0.76, 0.75, for the LR model were 0.77, 0.75. The DCA showed that the 3 machine learning models had good net benefits and clinical values in predicting EGFR mutation subtypes.Conclusion:18F-FDG PET/CT radiomics provide a non-invasive method for the identification of EGFR exon 19 deletion and exon 21 L858R missense mutations in patients with NSCLC, which may help the clinical decision-making and the formulation of individualized treatment plan.