Radiomics analysis to predict pulmonary nodule malignancy using machine learning approaches.

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Low-dose CT screening reduces lung cancer mortality but struggles to differentiate malignant from benign pulmonary nodules.
• Accurate identification of malignant nodules is crucial for effective lung cancer screening and management.
• Indeterminate nodules pose a significant challenge in baseline screening assessments.

Purpose of the Study:

• To develop and validate prediction models using radiological features to distinguish benign from malignant pulmonary lesions detected during baseline screening.
• To assess the performance of machine learning models in predicting the risk of malignancy for pulmonary nodules.

Main Methods:

• Extracted 2060 radiomic features from 16,797 nodules across four international lung cancer screening studies.
• Filtered features to 642 radiomic and 9 epidemiological variables for model development.
• Trained and evaluated three machine learning models (eXtreme Gradient Boosted Trees, random forest, LASSO) using cross-validation and grid search.

Main Results:

• The LASSO model achieved the highest predictive performance with a test-set Area Under the Curve (AUC) of 0.93.
• The developed radiomics model outperformed the Pan-Canadian Early Detection of Lung Cancer model (AUC 0.87).
• The model demonstrated high accuracy for both solid (AUC 0.93) and subsolid nodules (AUC 0.91).

Conclusions:

• Highly accurate machine learning models were developed using radiomic and epidemiological features.
• These models show potential for assessing indeterminate screen-detected pulmonary nodules for malignancy risk.
• The findings support the use of advanced computational methods in lung cancer screening interpretation.