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

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Radiology

Background:

• Mammography is vital for breast cancer screening but faces challenges like low contrast, noise, and artifacts.
• Accurate breast segmentation is essential for effective Computer-Aided Diagnosis (CAD) systems in mammography.
• Existing methods struggle with precise breast delineation and pectoral muscle exclusion in various mammogram views.

Purpose of the Study:

• To introduce and evaluate the SAM-breast model for automated breast segmentation in mammograms.
• To enhance the accuracy of breast region delineation and pectoral muscle exclusion.
• To assess the model's performance across diverse datasets and mammogram views (MLO and CC).

Main Methods:

• Adapted the Segment Anything Model (SAM) into the SAM-breast model for mammogram segmentation.
• Trained the model on a large, multi-center proprietary dataset comprising 2492 mammograms.
• Validated performance using independent test images from five datasets (two proprietary, three public).

Main Results:

• Achieved a high overall Dice Similarity Coefficient (DSC) of 99.22% ± 1.13 and Intersection over Union (IoU) of 98.48% ± 2.10.
• Demonstrated consistent performance across different datasets, vendors, and image resolutions.
• Outperformed baseline and other deep learning-based segmentation methods.

Conclusions:

• The SAM-breast model effectively segments the breast region in mammograms, showcasing SAM's adaptability.
• The method provides robust, flexible, and generalizable breast segmentation capabilities.
• This advancement holds significant potential for improving computer-aided diagnosis in breast cancer screening.