Radiomics analysis to predict pulmonary nodule malignancy using machine learning approaches

Purpose Screening with low-dose computed tomography can reduce lung cancer-related mortality. However, most screen-detected pulmonary abnormalities do not develop into cancer and it remains challenging to identify high-risk nodules among those with indeterminate appearance. We aim to develop and validate prediction models to discriminate between benign and malignant pulmonary lesions based on radiological features. Methods Using four international lung cancer screening studies, we extracted 2,060 radiomic features for each of 16,797 nodules among 6,865 participants. After filtering out redundant and low-quality radiomic features, 642 radiomic and 9 epidemiologic features remained for model development. We used cross-validation and grid search to assess three machine learning models (XGBoost, Random Forest, LASSO) for their ability to accurately predict risk of malignancy for pulmonary nodules. We fit the top-performing ML model in the full training set. We report model performance based on the area under the curve (AUC) and calibration metrics in the held-out test set. Results The ML models that yielded the best predictive performance in cross-validation were XGBoost and LASSO, and among these models, LASSO had superior model calibration, which we considered to be the optimal model. We fit the final LASSO model based on the optimized hyperparameter from cross-validation. Our radiomics model was both well-calibrated and had a test-set AUC of 0.930 (95% CI: 0.901-0.957) and out-performed the established Brock model (AUC=0.868, 95% CI: 0.847-0.888) for nodule assessment. Conclusion We developed highly-accurate machine learning models based on radiomic and epidemiologic features from four international lung cancer screening studies that may be suitable for assessing suspicious, but indeterminate, screen-detected pulmonary nodules for risk of malignancy.