Breast Delineation in Full-Field Digital Mammography Using the Segment Anything Model

• DOI: 10.3390/diagnostics14101015
• Publication Date: 2024-05-15T15:31:52Z
• Authors: Andrés Larroza; Francisco Javier ...; Raquel Tendero; Juan Carlos Perez...; Marta Román; Rafael Llobet
• Keywords: Breast ultrasound; Breast imaging; Robustness

Breast cancer is a major health concern worldwide. Mammography, cost-effective and accurate tool, crucial in combating this issue. However, low contrast, noise, artifacts can limit the diagnostic capabilities of radiologists. Computer-Aided Diagnosis (CAD) systems have been developed to overcome these challenges, with outlining breast being critical step for further analysis. This study introduces SAM-breast model, an adaptation Segment Anything Model (SAM) segmenting region mammograms. method enhances delineation exclusion pectoral muscle both medio lateral-oblique (MLO) cranio-caudal (CC) views. We trained models using large, multi-center proprietary dataset 2492 The proposed model achieved highest overall Dice Similarity Coefficient (DSC) 99.22% ± 1.13 Intersection over Union (IoU) 98.48% 2.10 independent test images from five different datasets (two three publicly available). results are consistent across datasets, regardless vendor or image resolution. Compared other baseline deep learning-based methods, exhibits enhanced performance. demonstrates power SAM adapt when it tailored specific tasks, case, Comprehensive evaluations diverse datasets—both private public—attest method’s robustness, flexibility, generalization capabilities.