Semi-Supervised Deep Learning-Based Model for Segmentation of Breast Arterial Calcification on Screening Mammograms.

Area of Science:

• Radiology and medical imaging
• Artificial intelligence in healthcare
• Cardiovascular disease risk stratification

Background:

• Breast arterial calcification (BAC) on mammograms is an independent cardiovascular disease (CVD) risk factor.
• Current BAC assessment lacks standardization and relies on subjective interpretation.
• Automated methods are needed for objective and reproducible BAC grading.

Purpose of the Study:

• To develop and validate a semi-supervised deep learning (DL) model for automated BAC severity grading.
• To enhance the model's generalizability across different mammography systems.
• To align automated grading with clinical consensus and improve CVD risk stratification.

Main Methods:

• A U-Net-based DL model was trained on annotated mammograms and enhanced with 6000 unlabeled images using progressive pseudo-labeling.
• BAC severity was graded by percentage area coverage, benchmarked against radiologist assessments using Canadian Society of Breast Imaging guidelines.
• Model performance was evaluated using Jaccard Similarity Coefficient, accuracy, precision, F1-score, recall, sensitivity, specificity, and area under the curve (AUC).

Main Results:

• The DL model achieved high performance with a Jaccard Similarity Coefficient of 0.614, accuracy of 0.991, and F1-score of 0.756.
• Excellent agreement with expert radiologists was observed (weighted kappa = 0.90).
• The model demonstrated strong performance in detecting clinically significant (Grade 3) BAC, with an AUC of 0.87, sensitivity of 0.80, and specificity of 0.93.

Conclusions:

• The developed semi-supervised DL framework offers a promising, standardized approach for automated BAC grading.
• This technology has the potential for clinical adoption within mammography workflows.
• Improved BAC assessment can enhance cardiovascular risk stratification for women.