Imaging phenotyping using 18F-FDG PET/CT radiomics to predict micropapillary and solid pattern in lung adenocarcinoma

Abstract

Abstract Objectives To develop and validate a machine learning model using 18F-FDG PET/CT radiomics signature and clinical features to predict the presence of micropapillary and solid (MP/S) components in lung adenocarcinoma. Methods Eight hundred and forty-six patients who underwent preoperative PET/CT with pathologically confirmed adenocarcinoma were enrolled. After segmentation, 1688 radiomics features were extracted from PET/CT and selected to construct predictive models. Then, we developed a nomogram based on PET/CT radiomics integrated with clinical features. Receiver operating curves, calibration curves, and decision curve analysis (DCA) were performed for diagnostics assessment and test of the developed models for distinguishing patients with MP/S components from the patients without. Results PET/CT radiomics-clinical combined model could well distinguish patients with MP/S components from those without MP/S components (AUC = 0.87), which performed better than PET (AUC = 0.829, p < 0.05) or CT (AUC = 0.827, p < 0.05) radiomics models in the training cohort. In test cohorts, radiomics-clinical combined model outperformed the PET radiomics model in test cohort 1 (AUC = 0.859 vs 0.799, p < 0.05) and the CT radiomics model in test cohort 2 (AUC = 0.880 vs 0.829, p < 0.05). Calibration curve indicated good coherence between all model prediction and the actual observation in training and test cohorts. DCA revealed PET/CT radiomics-clinical model exerted the highest clinical benefit. Conclusion 18F-FDG PET/CT radiomics signatures could achieve promising prediction efficiency to identify the presence of MP/S components in adenocarcinoma patients to help the clinician decide on personalized treatment and surveillance strategies. The PET/CT radiomics-clinical combined model performed best. Critical relevance statement 18F-FDG PET/CT radiomics signatures could achieve promising prediction efficiency to identify the presence of micropapillary and solid components in adenocarcinoma patients to help the clinician decide on personalized treatment and surveillance strategies. Graphical Abstract

Keywords

• Positron emission tomography
• Computed tomography
• Radiomics
• Lung adenocarcinoma
• Machine learning