A segment anything model-driven framework for multimodal active learning in breast cancer segmentation.

Medical imaging advancements have revolutionized clinical workflows; however, persistent challenges in annotation efficiency and cost continue to hinder scalable analysis. We propose a novel framework integrating the Segment Anything Model (SAM) with multimodal active learning to enhance breast cancer segmentation in Spectral Detector CT (SDCT) imaging. Our methodology combines SAM’s prompt-based segmentation capabilities with SDCT’s inherent dual-layer spectral consistency, enabling efficient annotation through expert-guided point prompts. Using a clinical dataset comprising 4996 slices from 162 breast cancer patients (32 test, 98 train, 32 validation), we demonstrate that multimodal integration of Iodine Density (IoD) and Z Effective (ZE) modalities achieves 96.75% of the performance level attained through expert manual annotations during model training, while reducing annotation costs by over 95%. Under intermediate annotation budgets, multi-click inputs and hybrid bounding box-click schemes consistently outperform single-click and box-only strategies, underscoring the value of richer spatial prompts for SAM. This SAM-driven active learning paradigm establishes a clinically viable pathway to maintain diagnostic-grade segmentation accuracy while significantly reducing resource-intensive manual annotation, thereby accelerating the translation of quantitative imaging biomarkers into routine breast cancer management.