Abstract
OBJECTIVE: This study aimed to identify CT-based radiomic alterations associated with radiation pneumonitis (RP) and to evaluate the feasibility of machine learning classifiers for personalized RP diagnosis in breast cancer patients using these radiomic signatures. METHODS: A total of 146 planning CT scans (pre- and post-radiotherapy) from 73 breast cancer patients with confirmed RP were retrospectively analyzed. The entire lung was delineated as the region of interest (ROI), and 1,834 radiomic features were extracted using PyRadiomics. Feature selection was performed sequentially using the Mann-Whitney U-test (p < 0.05), Spearman's rank correlation (|ρ| < 0.9), and least absolute shrinkage and selection operator (LASSO). Eight classifiers [logistic regression (LR), support vector machine (SVM), K-nearest neighbor (KNN), random forest (RF), Extra Tree (ET), XGBoost, LightGBM, and multilayer perceptron (MLP)] were trained and evaluated using accuracy, area under the receiver operating characteristic curve (AUC) with 95% confidence intervals, sensitivity, and specificity. RESULTS: In the independent test cohort, LR achieved the highest performance [accuracy 0.897, AUC 0.929 (95% CI, 0.838-1.000), sensitivity 0.786, and specificity 1.000]. LightGBM and MLP also exhibited robust discrimination with AUC values of 0.855 (95% CI, 0.719-0.990) and 0.848 (95% CI, 0.705-0.991), respectively. Five texture-oriented and four first-order features were retained, underscoring the importance of texture-focused extractors [wavelet and local binary pattern (LBP)]. CONCLUSION: CT-derived radiomic signatures combined with machine learning classifiers enable the accurate detection of RP in breast cancer patients. Texture-oriented feature selection enhances model discrimination, providing potential for the personalized diagnosis of RP in breast cancer patients and adaptive treatment planning.