Background: Breast cancer Mammography (MAM) screening was proven to improve survival worldwide. However, Mammography faces limitations in screening Asian women, who often have denser breasts. While ultrasound is a suitable alternative, its widespread use is hindered by dependence on operator skills. Our team developed a portable artificial intelligence (AI)-assisted breast ultrasound (US) system, trained on a large dataset of over 300,000 breast lesions, to overcome this barrier by assisting in standardized scan interpretation.
 
Objectives: To evaluate the clinical diagnostic accuracy of a portable AI-assisted US system and its effectiveness in a community-based screening program, aiming to provide scientific evidence for its implementation in large-scale population screening.
 
Methods: This project integrated clinical validation and community-based evaluation. First, a prospective diagnostic trial was conducted at Fudan University Shanghai Cancer Center. A total of 360 women underwent both AI-assisted US (performed by sonographers using a standardized scanning protocol) and conventional diagnostic US. Performance was compared for lesion detection and diagnostic accuracy. Building on these findings, a prospective, cluster-controlled, population-based trial was implemented within a national screening program in Shanghai. Two districts were assigned as clusters: one (n=8,736) received screening with the AI-assisted US system operated by trained healthcare personnel using the same standardized protocol, while the other (n=13,054) received routine ultrasound screening. Outcomes, including cancer detection and stage distribution, were assessed over a one-year follow-up.
 
Results: In the clinical setting, the AI-assisted US system demonstrated high diagnostic accuracy, with a sensitivity of 93.3% (95% CI: 80.7–98.3%) and specificity of 100% (99.5–100%) for detecting BI-RADS category 4 and above lesions, comparable to conventional ultrasound (table 3). For breast cancer diagnosis, both methods showed identical performance (sensitivity: 80.0%; specificity: 88.6%, table 4). In the community screening cohort, the AI-assisted screening group achieved a significantly higher screening sensitivity of 75.0% (54.8–88.6%) compared to 42.8% (22.6–65.6%) in the routine screening group (P<0.05). This translated to the detection of significantly more breast cancers (21 vs. 9, P=0.001) and a higher proportion of early-stage (0/I/II) cancers among screen-detected cases (95.2% vs. 88.9%, P<0.001) by the AI-assisted approach. The screen-positive rates and specificities were similar between the two community groups.
 
Conclusions: This project demonstrates that a portable AI-assisted ultrasound system is an effective tool for breast cancer detection. It performs comparably to expert sonography in a diagnostic clinical setting and, critically, demonstrates superior effectiveness in real-world community screening by significantly improving sensitivity and increasing the detection of early-stage cancers. The system mitigates the dependency on high-level operator expertise and shows strong potential to enhance the scalability and effectiveness of population-based breast cancer screening programs, particularly in regions with resource limitations. Further analysis of cost-effectiveness from the community trial is ongoing to support policy translation

Clinical Characteristics of AI and Routine US screen-detected and interval cancer by stage