Development and internal validation of a multimodal ultrasound and Radiomics‑Based nomogram for differentiating granulomatous lobular mastitis from invasive ductal carcinoma

Granulomatous lobular mastitis (GLM) and invasive ductal carcinoma (IDC) frequently present with overlapping ultrasound features, often leading to unnecessary biopsies. This study aimed to develop and internally validate a multimodal ultrasound (MMUS)-based nomogram for differentiating these two entities, and to evaluate the incremental value of adding a radiomics signature. Granulomatous lobular mastitis (GLM) and invasive ductal carcinoma (IDC) frequently present with overlapping ultrasound features, often leading to unnecessary biopsies. This study aimed to develop and internally validate a multimodal ultrasound (MMUS)-based nomogram for differentiating these two entities, and to evaluate the incremental value of adding a radiomics signature. A total of 460 patients with pathologically confirmed GLM or IDC were retrospectively enrolled and randomly divided into a training cohort ( n = 322) and an internal-validation cohort ( n = 138) at a 7:3 ratio. Radiomic features were extracted from grayscale ultrasound images, and a radiomics score (Radscore) was constructed using LASSO regression. Multivariate logistic regression was performed to identify the final set of predictors. Four models were compared: an age-only model, the Clinical MMUS model, the radiomics-only model, and the combined model. Model performance was assessed using AUC, calibration (intercept, slope, Brier score), and decision curve analysis (DCA). Six predictors retained in the final multivariable model—age, shape, margin, calcifications, elasticity score, and Radscore—were integrated into the nomogram. In the validation cohort, the Clinical MMUS model achieved an AUC of 0.978 (95% CI: 0.956–0.999), sensitivity of 0.972 (95% CI: 0.921–0.994), and specificity of 0.867 (95% CI: 0.693–0.962). The combined model demonstrated a sensitivity–specificity redistribution: AUC of 0.985 (95% CI: 0.967–1.000), sensitivity of 0.907 (95% CI: 0.836–0.955), and specificity of 0.967 (95% CI: 0.828–0.999). The difference in AUC between the MMUS and combined models was not statistically significant (ΔAUC = −0.0070, p = 0.153). In the validation cohort, the age-only model achieved an AUC of 0.884 (95% CI: 0.813–0.954), and the Clinical MMUS model significantly outperformed this age-only baseline (ΔAUC = 0.094, p = 0.005). Calibration analysis showed acceptable agreement in the internal-validation cohort for both models. DCA using raw net benefit showed largely overlapping curves for the MMUS and combined models across the exploratory threshold range of 0.1–0.5, without consistent incremental benefit from radiomics. Notably, the Clinical MMUS model, which excludes the radiomics signature, demonstrated strong discriminative performance with a validation AUC of 0.978. The MMUS-based nomogram demonstrated strong discriminative performance for differentiating GLM from IDC. The addition of the radiomics signature provided no statistically significant or clinically meaningful incremental benefit. This model may serve as a quantitative decision-support tool, though it was developed in a pathology-enriched binary cohort limited to path-proven GLM or IDC and cannot be generalized to unselected breast lesions. Its potential utility in cases with equivocal conventional ultrasound findings remains a hypothesis for future research. External validation is required before clinical adoption.