18F-FDG PET/CT-based radiomics model for predicting the degree of pathological differentiation in non-small cell lung cancer: a multicentre study

•Few reports on PET/CT radiomics to predict pathological differentiation of lung cancer. •We explored a non-invasive method to predict pathological differentiation in NSCLC. •Multicenter study may contribute to the robustness and generalizability of our model. AIM To explore the value of 2-[18F]-fluoro-2-deoxy-d-glucose (FDG) positron-emission tomography (PET)/computed tomography (CT)-based radiomics model for predicting the degree of pathological differentiation in non-small-cell lung cancer (NSCLC). MATERIALS AND METHODS Clinical characteristics of 182 NSCLC patients from four centres were collected, and radiomics features were extracted from 18F-FDG PET/CT images. Three logistic regression prediction models were established: clinical model; radiomics model; and nomogram combining radiomics signatures and clinical features. The predictive ability of the models was assessed using receiver operating characteristics curve analysis. RESULTS Patients from centre 1 were assigned randomly to the training and internal validation cohorts (7:3 ratio); patients from centres 2–4 served as the external validation cohort. The area under the curve (AUC) values for the clinical model in the training, internal validation, and external validation cohort were 0.74 (95% confidence interval [CI] = 0.64–0.84), 0.64 (95% CI = 0.46–0.81), and 0.74 (95% CI = 0.60–0.88), respectively. In the training (AUC: 0.84 [95% CI = 0.77–0.92]), internal validation (AUC: 0.81 [95% CI = 0.67–0.95]), and external validation cohorts (AUC: 0.74 [95% CI = 0.58–0.89]), the radiomics model showed good predictive ability for differentiation. Compared to the clinical and radiomics models, the nomogram has relatively better diagnostic performance, and the AUC values for nomogram in the training, internal validation, and external validation cohort were 0.86 (95% CI = 0.78–0.93), 0.83 (95% CI = 0.70–0.96), and 0.77 (95% CI = 0.62–0.92), respectively. CONCLUSIONS The 18F-FDG PET/CT-based radiomics model showed good ability for predicting the degree of differentiation of NSCLC. The nomogram combining the radiomics signature and clinical features has relatively better diagnostic performance. To explore the value of 2-[18F]-fluoro-2-deoxy-d-glucose (FDG) positron-emission tomography (PET)/computed tomography (CT)-based radiomics model for predicting the degree of pathological differentiation in non-small-cell lung cancer (NSCLC). Clinical characteristics of 182 NSCLC patients from four centres were collected, and radiomics features were extracted from 18F-FDG PET/CT images. Three logistic regression prediction models were established: clinical model; radiomics model; and nomogram combining radiomics signatures and clinical features. The predictive ability of the models was assessed using receiver operating characteristics curve analysis. Patients from centre 1 were assigned randomly to the training and internal validation cohorts (7:3 ratio); patients from centres 2–4 served as the external validation cohort. The area under the curve (AUC) values for the clinical model in the training, internal validation, and external validation cohort were 0.74 (95% confidence interval [CI] = 0.64–0.84), 0.64 (95% CI = 0.46–0.81), and 0.74 (95% CI = 0.60–0.88), respectively. In the training (AUC: 0.84 [95% CI = 0.77–0.92]), internal validation (AUC: 0.81 [95% CI = 0.67–0.95]), and external validation cohorts (AUC: 0.74 [95% CI = 0.58–0.89]), the radiomics model showed good predictive ability for differentiation. Compared to the clinical and radiomics models, the nomogram has relatively better diagnostic performance, and the AUC values for nomogram in the training, internal validation, and external validation cohort were 0.86 (95% CI = 0.78–0.93), 0.83 (95% CI = 0.70–0.96), and 0.77 (95% CI = 0.62–0.92), respectively. The 18F-FDG PET/CT-based radiomics model showed good ability for predicting the degree of differentiation of NSCLC. The nomogram combining the radiomics signature and clinical features has relatively better diagnostic performance.