Artificial intelligence-based multimodal integration of ultrasound and digital breast tomosynthesis for breast-level risk classification.
Authors
Affiliations (10)
Affiliations (10)
- Guangdong Provincial Key Laboratory of Malignant Tumor Epigenetics and Gene Regulation, Sun Yat-sen Memorial Hospital, Sun Yat-sen University, Guangzhou 510000, China.
- Guangdong-Hong Kong Joint Laboratory for RNA Medicine, Sun Yat-sen University, Guangzhou 510000, China.
- Department of Medical Oncology, Sun Yat-sen Memorial Hospital, Sun Yat-sen University, Guangzhou 510000, China.
- Breast Tumor Center, Sun Yat-sen Memorial Hospital, Sun Yat-sen University, Guangzhou 510000, China.
- Phase I Clinical Trial Centre, Sun Yat-sen Memorial Hospital, Sun Yat-sen University, Guangzhou 510000, China.
- Faculty of Medicine, Macau University of Science and Technology, Macau 999078, China.
- Faculty of Medicine and Artificial Intelligence Cross Disciplinary Research Institute, Macau University of Science and Technology, Macau 999078, China.
- Department of Medical Oncology, Liuzhou People's Hospital, Guangxi Medical University, Liuzhou 545006, China.
- Hui Ya Hospital of the First Affiliated Hospital, Sun Yat-Sen University, Huizhou 516081, China.
- Cellsvision (Guangzhou) Medical Technology Inc., Guangzhou 510000, China.
Abstract
Breast cancer imaging frequently combines ultrasound (US), digital mammography (DM), and digital breast tomosynthesis (DBT), yet integrating complementary findings remains labor-intensive. We developed a parallel-branch deep learning framework for breast-level risk classification from paired US, DM, and DBT examinations. The models were trained on 2187 breasts and evaluated in an internal validation cohort of 632 breasts and an independent pathology-confirmed cohort of 500 breasts. Six single- and dual-modality models were compared. In the internal validation cohort, US-DBT achieved the highest observed area under the curve (AUC) of 0.944 (95% CI, 0.926-0.963), exceeding US-DM and DM-DBT but not US; its sensitivity was 0.860 (95% CI, 0.805-0.904) and specificity was 0.904 (95% CI, 0.871-0.930). In the pathology-confirmed cohort, US-DBT achieved an AUC of 0.934 (95% CI, 0.913-0.955), exceeding US and DM-DBT but not US-DM. Its specificity was higher than that of all three models (0.955; 95% CI, 0.927-0.975; all adjusted <i>P </i>< 0.001), with a positive predictive value (PPV) of 0.958 (95% CI, 0.931-0.977) and sensitivity of 0.850 (95% CI, 0.807-0.887), which did not differ significantly from any of the three models. Performance remained favorable in dense breasts, lesions <2 cm, and lower-suspicion Breast Imaging Reporting and Data System (BI-RADS) strata. These findings identify improved specificity as the principal added value of US-DBT and support its potential use as an adjunctive breast-level tool for refining positive imaging findings and prioritizing further diagnostic evaluation. Prospective validation in representative screening populations is required.