Automating Medical SAM Adapter for Segmentation of Digital Breast Tomosynthesis Images

Sammanfattning

Accurate segmentation of breast lesions in digital breast tomosynthesis (DBT) remains a critical challenge for automated breast cancer screening. This study evaluates the performance of the state-of-the-art nnU-Net segmentation framework, the Medical SAM Adapter (Med-SA), and a combined nnU-Net–Med-SA pipeline on a curated DBT dataset. Manual annotations were performed on 285 cases, which showed great variability in lesion presentation and breast tissue density. Experimental results demonstrate that nnU-Net consistently outperforms Med-SA across all metrics, while Med-SA struggles despite prompt guidance. The combined nnU-Net–Med-SA pipeline did not improve segmentation quality over nnU-Net alone, indicating that sequential application of volumetric and prompt-based methods is insufficient for DBT data. These findings establish baseline performance for automated DBT segmentation and underscore the need for volumetric-aware adaptation of prompt-based models. Future work should explore improved volumetric prompting strategies, incorporation of BI-RADS information, and multi-modal imaging to enhance clinical applicability.