Optimizing Diagnosis of Non-Microcalcified BI-RADS 4 Breast Lesions Using Multimodal Ultrasound Logistic Regression.

To develop a logistic prediction model based on multimodal ultrasound to enhance the diagnostic performance for non-microcalcified BI-RADS 4 breast lesions. We retrospectively analyzed ultrasound data from 334 patients with BI-RADS 4 breast lesions, incorporating 10 multimodal features. Model 1 was constructed using the entire cohort, while Model 2 focused on a subset of 225 non-microcalcified cases, with features selected via Lasso regularization and performance evaluated through 10-fold cross-validation. Model 1 identified lesion size >2 cm (OR = 1.65, p = .041), microcalcification (OR = 3.62, p < .001), and Emax (OR = 1.02, p = .001) as independent predictors, with an AUC of 0.85 (95%CI: 0.78-0.91). Model 2 selected lesion size, Adler grade, and Emax as significant features, achieving an AUC of 0.88 (95%CI: 0.81-0.92), with a 10-fold cross-validated accuracy of 0.81, Kappa of .57, and Hosmer-Lemeshow test (χ2 = 5.23, p = .850) for calibration. The multimodal ultrasound-based logistic model significantly improves the diagnosis of non-microcalcified BI-RADS 4 breast lesions (AUC = 0.88), with lesion size, Adler grade, and Emax as key predictors, offering a cost-effective tool to reduce unnecessary biopsies in clinical practice.