Potential feature exploration and model development based on 18F-FDG PET/CT images for differentiating benign and malignant lung lesions

Purpose: The study is to explore potential features and develop classification models for distinguishing benign and malignant lung lesions based on CT-radiomics features and PET metabolic parameters extracted from PET/CT images. Materials and methods: A retrospective study was conducted in baseline 18F-flurodeoxyglucose positron emission tomography/ computed tomography (18F-FDG PET/CT) images of 135 patients. The dataset was utilized for feature extraction of CT-radiomics features and PET metabolic parameters based on volume of interest, then went through feature selection and model development with strategy of five-fold cross-validation. Specifically, model development used support vector machine, PET metabolic parameters selection used Akaike's information criterion, and CT-radiomics were reduced by the least absolute shrinkage and selection operator method then forward selection approach. The diagnostic performances of CT-radiomics, PET metabolic parameters and combination of both were illustrated by receiver operating characteristic (ROC) curves, and compared by Delong test. Five groups of selected PET metabolic parameters and CT-radiomics were counted, and potential features were found and analyzed with Mann-Whitney U test. Results: The CT-radiomics, PET metabolic parameters, and combination of both among five subsets showed mean area under the curve (AUC) of 0.820 +/- 0.053, 0.874 +/- 0.081, and 0.887 +/- 0.046, respectively. No significant differences in ROC among models were observed through pairwise comparison in each fold (P-value from 0.09 to 0.81, Delong test). The potential features were found to be SurfaceVolumeRatio and SUVpeak (P < 0.001 of both, U test). Conclusion: The classification models developed by CT-radiomics features and PET metabolic parameters based on PET/CT images have substantial diagnostic capacity on lung lesions.