An Effective Malignancy Prediction Model for Incidentally Detected Pulmonary Subsolid Nodules Based on Current and...

Area of Science:

• Pulmonary medicine
• Radiology
• Artificial Intelligence in Medicine

Background:

• Diagnosing and managing pulmonary subsolid nodules is difficult due to their slow growth and varied characteristics.
• Accurate prediction of malignancy is crucial for appropriate patient management.

Purpose of the Study:

• To develop and validate a decision tree-based model for predicting the malignancy of pulmonary subsolid nodules.
• To integrate radiomics features and their temporal changes into a predictive model.

Main Methods:

• A dataset of 2947 subsolid nodules was used for training, with independent test sets from outpatient scans (280 nodules) and the National Lung Cancer Screening Trial (NLST) (5171 nodules).
• A Computer-Aided Diagnosis system extracted 28 radiomics features; feature change rates were calculated between scans.
• XGBoost classification models were built and optimized using 5-fold cross-validation.

Main Results:

• The combined model using radiomics features and their change rates achieved high predictive performance.
• Areas Under the Curve (AUCs) were 0.977 on the outpatient test set and 0.955 on the NLST test set.
• The model demonstrated consistent performance across nodules of varying sizes, solid components, and CT scan intervals.

Conclusions:

• The developed decision tree model shows strong potential for predicting pulmonary subsolid nodule malignancy.
• This model can enhance diagnostic accuracy and inform follow-up strategies for clinicians.
• Integrating radiomics and temporal analysis offers a promising approach for subsolid nodule assessment.