Deep-learning model for background parenchymal enhancement classification in contrast-enhanced mammography

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Breast Cancer Diagnostics

Background:

• Breast background parenchymal enhancement (BPE) is a known risk factor for breast cancer.
• Current BPE assessment in contrast-enhanced mammography (CEM) relies on subjective radiologist interpretation (BI-RADS categories), leading to variability.
• Automated BPE classification methods exist for MRI but are lacking for CEM.

Purpose of the Study:

• To develop and evaluate a deep learning-based tool for automated BPE level classification in CEM.
• To assess the tool's performance, robustness to lesions, and compare it with existing methods.

Main Methods:

• A deep learning model was trained on 7012 CEM image pairs (low-energy and recombined).
• Model optimization involved analyzing image resolution, backbone architecture, and loss functions.
• Performance was evaluated using 4-class balanced accuracy and mean absolute error on a 1013 image pair dataset, including analysis of lesion presence.

Main Results:

• The optimized model achieved 71.5% 4-class balanced accuracy, with 98.8% of errors between adjacent classes.
• Binary classification (minimal/mild vs. moderate/marked) reached 93.0% accuracy.
• Model performance showed robustness to the presence of lesions, with only a slight, non-significant decrease in accuracy.

Conclusions:

• The developed deep learning tool provides accurate and repeatable BPE classification for CEM.
• This automated approach has the potential to overcome the limitations of subjective human assessment.
• The tool's performance is comparable to methods reported for CE-MRI, marking a significant advancement for CEM.