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

OBJECTIVE

To evaluate machine learning models based on HRCT radiomic features for distinguishing breast and colorectal cancer pulmonary metastases, and interpret the optimal model to aid clinical decision-making.

METHODS

This retrospective study enrolled 85 patients with pathologically confirmed pulmonary metastases. After radiomic feature extraction, the cohort was divided into a training set (n=59) and an independent test set (n=26) at a 7:3 ratio via stratified sampling. Data were processed with Z-score normalization, variance thresholding and PCA (45 principal components). Five classifiers were constructed: LR, linear SVM, RF, XGBoost and LightGBM. Model stability and performance were assessed by 5-fold stratified cross-validation and independent test validation.

RESULTS

The 45 principal components accounted for 99.92% of cumulative variance. LR showed optimal performance, with a test AUC of 0.9821, classification accuracy of 84.62%, and a mean cross-validation AUC of 0.9606 (95% CI: 0.8887-0.9895). The small training-test AUC difference (0.0179) indicated no severe overfitting. SVM ranked second (test AUC = 0.9405), while XGBoost and RF exhibited significant overfitting and LightGBM underfitting. The model's decision relied on key texture features; only GLSZM non-uniformity differed significantly between groups, consistent with their pathophysiological characteristics.

CONCLUSION

The PCA-reduced and regularization-optimized LR model has excellent generalization, stability and clinical interpretability in differentiating pulmonary metastases from breast and colorectal cancers. This study preliminarily highlights the potential of LR in high-dimensional small-sample scenarios, and provides a foundational methodological reference for future large-scale multicenter diagnostic research.