Machine learning-based photoacoustic/ultrasound radiomics models for differentiating benign and malignant breast lesions

This study evaluated the efficacy of the photoacoustic/ultrasound (PA/US) radiomics model in distinguishing benign and malignant breast lesions. Data from 348 patients (278 training, 70 testing) were analyzed. A total of 107 PA imaging features were extracted and subsequently refined to 39 non-zero features using a feature selection that applied Spearman correlation analysis, LASSO regression, and other feature dimensionality reduction methods. The K-nearest neighbor (KNN) algorithm developed the PA radiomics model, while logistic regression identified clinical variables for a clinical model. A combined model integrating PA radiomics and clinical features was created, and a grayscale ultrasound (US) radiomics model served as a control. ROC analysis showed AUCs of 0.830 (PA radiomics), 0.892 (clinical), and 0.915 (combined) in testing set, with the combined model achieving the highest discrimination (training AUC = 0.977; testing AUC = 0.915). The PA radiomics model outperformed the grayscale US model (AUC = 0.830 vs. 0.793) in the testing set. Decision curve analysis (DCA) confirmed that the combined model demonstrated a more favorable net benefit profile across a wide range of thresholds in this cohort. The combined PA radiomics and clinical model shows promising potential in differentiating benign and malignant breast lesions and could serve as a valuable non-invasive adjunctive diagnostic tool. Notably, PA radiomics demonstrated a statistically significant higher AUC than grayscale US radiomics, though its specificity requires further optimization.