AgentDS-BUS: a fine-tuning-free agentic breast ultrasound malignancy classification framework with decoupled segmentation and feature analysis

Breast ultrasound (BUS) malignancy assessment is challenged by acquisition variability and cross-dataset shift. We developed AgentDS-BUS, an agentic clinical decision-support framework that uses pretrained foundation models without additional BUS-specific model fine-tuning and decomposes malignancy assessment into lesion localization and segmentation, region-of-interest (ROI)-conditioned extraction of Breast Imaging Reporting and Data System (BI-RADS)-aligned evidence, and structured decision aggregation. In this implementation, lesion masks are obtained using consensus-gated Segment Anything Model 3 (SAM3), with clinician-provided bounding boxes used for fallback segmentation in low-consensus cases. Across 4,331 images from seven public BUS benchmarks and three vision-language model (VLM) backbones, AgentDS-BUS improved pooled AUROC/AUPRC over image-only prompting. Under Qwen3-VL, performance increased from 0.636/0.524 with image-only prompting to 0.717/0.606 for AgentDS-BUS with consensus-gated ROIs and 0.791/0.654 for AgentDS-BUS with ground-truth (GT) oracle ROIs. With Lingshu, AgentDS-BUS achieved 0.760/0.678 in the consensus-gated ROI setting and 0.807/0.715 in the GT-ROI setting. These findings support structured agentic decomposition and ROI-conditioned evidence extraction as a transparent approach to BUS decision support.