<p>This study presents an efficient sample annotation framework based on active learning, designed to enhance annotation quality while reducing both the volume of annotated samples and annotation time. The framework integrates an initial screening of high-value samples with an iterative selection mechanism that leverages measures of representativeness, diversity, and uncertainty. To further optimize the process, an automated annotation module is developed using the Segment Anything Model, renowned for its exceptional and consistent performance across diverse imaging modalities. A novel initialization strategy is proposed for sample selection, integrating multimodal medical imaging data to effectively identify high-value samples containing critical lesion information. Experimental results demonstrate that the proposed framework achieves 97.6% of the performance of fully annotated methods while requiring annotation for only 2.7% of breast cancer imaging samples and shortening annotation time. Moreover, as the number of annotated samples increases, the proposed method consistently outperforms traditional techniques in both accuracy and robustness. By significantly reducing the labor, time, and resources required for data annotation, this framework offers a constructive approach for advancing medical imaging applications in clinical practice.</p>