New multimodal AI model improves breast cancer screening accuracy

Breast cancer screening must balance diagnostic accuracy with workflow throughput. DM remains the traditional cornerstone, but its two-dimensional projection creates tissue-overlap noise, particularly in dense breasts, leading to false positives and obscured lesions. Ultrasound (US) is widely used as an adjunct because it is sensitive to soft-tissue masses in dense parenchyma, yet it is operator-dependent and less sensitive to microcalcifications. DBT improves lesion visibility and reduces recall, but generates large volumetric datasets that increase reading time and cognitive workload. AI has shown strong performance within single modalities, but most models cannot cross-verify complementary evidence across imaging streams. Given these challenges, in-depth research is needed on multimodal AI frameworks that integrate US, DM, and DBT for breast-level risk triage.

Researchers at Sun Yat-sen Memorial Hospital, Sun Yat-sen University, Guangzhou, China, and collaborating institutions report the findings (DOI: 10.1093/pcmedi/pbag023) in Precision Clinical Medicine in 2026 (Volume 9, Issue 3). The team developed and validated a parallel-branch deep learning framework for breast-level risk classification using paired US, DM, and DBT examinations. Models were trained on 2,187 breasts and evaluated in an internal validation cohort of 632 breasts and an independent pathology-confirmed cohort of 500 breasts, with histopathology serving as the reference standard. The study compared six prespecified model configurations.

The study compared three single-modality models—US, DM, and DBT—and three dual-modality models: US-DM, DM-DBT, and US-DBT. The US-DBT model achieved the highest observed area under the receiver operating characteristic curve (AUC) in both validation cohorts: 0.944 (95% confidence interval [CI], 0.926-0.963) internally and 0.934 (95% CI, 0.913-0.955) in the pathology-confirmed cohort. In the pathology-confirmed cohort, its specificity reached 0.955 (95% CI, 0.927-0.975), and its positive predictive value (PPV) was 0.958 (95% CI, 0.931-0.977), while sensitivity was 0.850 (95% CI, 0.807-0.887) and did not differ significantly from the main comparator models. Performance remained favorable in dense breasts, lesions smaller than 2 cm, and lower-suspicion Breast Imaging Reporting and Data System (BI-RADS) strata. The architecture used modality-specific branches, a modified 2.5D ResNet18 with grouped convolutions for DBT, a Convolutional Block Attention Module (CBAM), feature-level fusion, and a multilayer perceptron (MLP) classifier. Each model produced a breast-level probability score using a prespecified 0.50 threshold. The approach is designed for breast-level, not participant-level, output.

The authors said the findings point to a practical role for complementary imaging. They said US contributes lesion-level information such as echogenicity, margins, posterior acoustic features, and internal structure, while DBT captures structural distortion, spiculation, and microcalcifications. Together, these data streams may help radiologists refine positive or discordant breast imaging findings. They emphasized that the model is not intended as a stand-alone screening or diagnostic system and that prospective, multicenter validation is needed before clinical implementation. The most consistent benefit, they said, was improved specificity without a significant loss of sensitivity.

Clinically, a US-DBT adjunct could help prioritize cases that warrant further diagnostic evaluation and reduce unnecessary escalation caused by false-positive findings, especially in dense breasts, small lesions, and lower-suspicion BI-RADS categories. By converting complementary sonographic and tomosynthesis information into a unified breast-level risk estimate, the approach may support radiologist-led triage and more selective work-up. However, the authors caution that the retrospective, single-center design limits generalizability. Broader implementation will depend on external validation across institutions, imaging platforms, and patient populations, as well as prospective testing of real-time screening workflows and clinician trust. If validated, such a tool could complement, rather than replace, routine assessment and reduce avoidable anxiety.