Evaluation of radiomic feature harmonization techniques for benign and malignant pulmonary nodules

Radiomic features derived from pulmonary nodule CT images may vary due to acquisitional factors, such as scanner settings, as well as biological factors, such as tissue type. Using a dataset of over 500 chest CT scans, we evaluated techniques to harmonize radiomic features in both benign and malignant nodules. We found that harmonizing distinct tissue types either separately or with a biological covariate was necessary to eliminate acquisitional variability in these features. A predictive model trained using our harmonized dataset was able to process routine lung cancer screening scans and identify nodules that later developed malignancy.