Perinodular Parenchymal Features Improve Indeterminate Lung Nodule Classification.

Area of Science:

• Radiology and Medical Imaging
• Oncology
• Data Science in Medicine

Background:

• Radiomics extracts quantitative imaging features for non-invasive assessment of pulmonary nodules.
• Distinguishing malignant from benign nodules is crucial for patient management.
• Low-dose computed tomography (LDCT) is widely used for lung cancer screening.

Purpose of the Study:

• To evaluate the effectiveness of perinodular radiomics in identifying malignant pulmonary nodules.
• To compare the diagnostic performance of radiomic features from the tumor itself versus surrounding lung parenchyma.
• To assess combined radiomic features for improved malignancy prediction.

Main Methods:

• Utilized data from the National Lung Screening Trial (NLST) for nodules 4-20mm.
• Segmented nodules into Tumor, 10mm Band, 15mm Band, and Tumor Size datasets.
• Employed Support-Vector Machine (SVM), Random Forest (RF), and LASSO models for malignancy prediction.
• Performed rigorous cross-validation (10-fold with 10 repetitions) for performance evaluation.

Main Results:

• Random Forest models achieved AUCs of 84.44% (Tumor), 84.09% (10mm Band), and 81.57% (15mm Band).
• Combined datasets (10mm Band + Tumor, 15mm Band + Tumor) significantly outperformed the Tumor dataset (AUCs 87.87%, 86.75% vs. 84.44%).
• The 10mm Band + Tumor dataset achieved high AUCs with SVM (84.71%) and LASSO (88.91%).

Conclusions:

• Combining perinodular radiomic features with tumor-specific features significantly enhances the differentiation between benign and malignant lung nodules.
• Parenchymal radiomic features capture valuable diagnostic information not present within the nodule alone.
• This integrated radiomic approach offers improved accuracy for non-invasive lung nodule characterization.