Integrating Ultrasound and DCE-MRI Improves Accuracy in Differentiating Breast Adenosis from Carcinoma

Discriminating breast adenosis from carcinoma remains challenging due to their overlapping imaging appearances. This study aimed to develop and validate a fusion model combining ultrasound (US) and MRI-based radiomics for improved differentiation.

In this retrospective study, 260 patients (147 adenosis, 113 carcinoma) with preoperative US and MRI from March 2019 to July 2025 were enrolled and randomly split into training (n=182) and testing (n=78) cohorts at a 7:3 ratio. Radiomic features were extracted from grayscale US images depicting the largest lesion diameter and from the second-phase enhancement of DCE-MRI. After selection via maximum relevance minimum redundancy(mRMR) and Least Absolute Shrinkage and Selection Operator (LASSO) regression, radiomics models (US_Rad, MRI_Rad) were built based on selected features using the optimal algorithm among 12 candidates. A BI-RADS model based on significant US and MRI BI-RADS features was also constructed. A fusion model integrated US_Rad, MRI_Rad, BI-RADS, and age. Receiver operating characteristic (ROC) curves were used to evaluate the predictive performance of the four models.

In the training cohort, the area under the curve (AUC) for BI-RADS and radiomic models based on US (US_Rad) and MRI (MRI_Rad) were 0.89 (95% CI, 0.84-0.93), 0.84 (95% CI, 0.78-0.90), and 0.85 (95% CI, 0.79-0.90), respectively. In the testing cohort, the AUCs were 0.91 (95% CI, 0.83-0.97), 0.82 (95% CI, 0.73-0.91), and 0.82 (95% CI, 0.73-0.90). The fusion model achieved superior AUCs of 0.96 (95% CI, 0.94-0.98, training) and 0.97 (95% CI, 0.92-1.00, testing), significantly outperforming both the unimodal radiomic model and the BI-RADS model (all p < 0.01).

Our findings suggest that the US-MRI radiomics fusion model exhibits potential in distinguishing breast adenosis from carcinoma, which may help reduce unnecessary surgeries. However, further multi-center validation is required to evaluate its generalizability before clinical application.