Radiomics-based high-resolution CT analysis for differentiating primary tumor sources of pulmonary metastases

Area of Science:

• Radiology and Medical Imaging
• Machine Learning in Oncology
• Computational Pathology

Background:

• Pulmonary metastases from breast and colorectal cancers present diagnostic challenges.
• High-resolution computed tomography (HRCT) offers rich textural information for analysis.
• Machine learning (ML) can potentially enhance the differentiation of metastatic origins.

Purpose of the Study:

• To evaluate ML models for distinguishing pulmonary metastases from breast and colorectal cancers using HRCT radiomic features.
• To identify the optimal model for clinical decision-making support.

Main Methods:

• Retrospective analysis of 85 patients with confirmed pulmonary metastases.
• Radiomic feature extraction followed by Z-score normalization, variance thresholding, and principal component analysis (PCA).
• Development and validation of five classifiers (LR, SVM, RF, XGBoost, LightGBM) using cross-validation and an independent test set.

Main Results:

• The logistic regression (LR) model achieved the highest performance with a test AUC of 0.9821 and 84.62% accuracy.
• A small difference between training and test AUC indicated no significant overfitting.
• Key texture features, particularly GLSZM non-uniformity, were significant differentiators, aligning with pathophysiological characteristics.

Conclusions:

• The PCA-reduced and regularization-optimized LR model shows excellent generalization, stability, and interpretability for differentiating pulmonary metastases.
• Logistic regression shows promise in high-dimensional, small-sample scenarios for cancer diagnosis.
• This study provides a methodological foundation for future large-scale multicenter diagnostic research.