Radiomics Analysis for Predicting Growth of Subsolid Lung Nodules on CT

Background Accurate identification of growing subsolid nodules is crucial for effective risk stratification and the early detection of invasive lung cancer, allowing for timely treatment while avoiding unnecessary surgery on low-risk nodules that would otherwise remain stable. The traditional method of risk stratification, which relies on qualitative visual analysis of CT scans, remains challenging. Therefore, this study aims to leverage a longitudinal dataset of subsolid nodules on CT and develop radiomic and clinical feature-based models to identify nodules that are likely to grow over time.

Purpose The purpose of this study is to develop a machine learning model to predict the growth of subsolid nodules using a combination of radiomic and clinical features.

Materials and Methods A retrospective study was conducted on a cohort of patients who had undergone chest CT scans at a single institution between 2015 and 2019. Corresponding radiology reports were used to extract ground truth labels for the nodules’ growth status (i.e., increased, decreased, unchanged), as well as other related information such as nodule size, nodule type, location, slice location, component sizes, and compared studies. An automated NLP pipeline was employed for the extraction of labels. Additionally, associated CT scans were processed through a commercial nodule characterization algorithm, which generated parameters such as nodule size, location, and 3D segmentation coordinates. Utilizing this dataset, along with radiomic features computed from pyRadiomics and clinical features (e.g., patient age and sex), models for predicting the growth of subsolid nodules were developed. The primary metric used to evaluate model performance was the AUC-ROC, assessed on both the independent validation set and averaged across 5-fold cross-validation. A total of 15 features were used for the model, selected through a combination of recursive feature elimination, random forest importance analysis, and univariate selection during cross-validation.

Results The final dataset comprises 950 subsolid nodules, each with corresponding growth statuses and 3D segmentations (799 unchanged, 139 growing, 12 decreasing). Among various models, support vector machine (SVM) achieved the highest AUC of 0.81 on both 5-fold cross-validation and the independent validation set. Furthermore, in the statistical analysis of the association between radiomic and clinical features and increasing versus unchanged nodules, 88 radiomic features were identified as statistically significant (p < 0.05) through a Mann-Whitney U test, with 82 of these radiomic features being highly statistically significant (p < 0.01). Notable radiomic features, including Voxel Volume, Run Length Non-Uniformity, and Dependence Non-Uniformity, demonstrated high feature importance in identifying growing nodules in the SVM model. Among the clinical features, Patient Age exhibited high feature importance.

Conclusion A model based on combined radiomic and clinical features was trained to predict subsolid nodules that would grow over time. Dependence Non-Uniformity, Run Length Non-Uniformity, Voxel Volume, Gray Level Non-Uniformity, and Patient Age were among the most predictive features for identifying high-risk, growing nodules.