A deep learning system to obtain the optimal parameters for a threshold-based breast and dense tissue segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast cancer is the most common cancer in women, necessitating early detection through screening programs.
• Breast density on mammograms is a key risk factor for breast cancer development.
• Automating breast density assessment is crucial due to increasing mammogram volumes and variations in imaging devices.

Purpose of the Study:

• To develop a fully automated deep learning framework for reliable breast density estimation from digital mammograms.
• To address challenges in automated segmentation, including inter-device image variability and the absence of a gold standard.
• To segment fibroglandular tissue accurately, including breast detection and pectoral muscle exclusion.

Main Methods:

• A multi-center study involving 1785 women and 6680 mammograms for training and testing.
• Implementation of histogram normalization to standardize image acquisition differences.
• Utilized a regression architecture with a DICE score-based loss function for segmentation.

Main Results:

• The automated framework achieved a DICE score of 0.77, comparable to inter-radiologist concordance (0.77).
• For high-quality images, the framework achieved a DICE score of 0.84, versus 0.76 for radiologists.
• The system demonstrated robust performance across different imaging devices.

Conclusions:

• A deep learning-based automatic breast density estimator shows performance comparable to experienced radiologists.
• This automated system has the potential to assist radiologists, improving efficiency in breast cancer screening.
• The framework offers a reliable solution for breast density assessment in large-scale screening programs.