Self-Supervised Contrastive Learning With Attention Fusion for Enhanced Breast Cancer Diagnosis From Mammography

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Mammography interpretation requires integrating information from multiple views (craniocaudal, mediolateral oblique).
• Accurate detection of subtle abnormalities like calcifications is crucial for early cancer diagnosis.
• Current AI models may struggle with view-invariance and contralateral asymmetry.

Purpose of the Study:

• To develop an anatomy-aware AI framework for improved mammography interpretation.
• To enhance the model's ability to reconcile evidence across different breast views.
• To improve the performance of AI in screening mammography, particularly in high-specificity scenarios.

Main Methods:

• Proposed a self-supervised anatomy-aware framework with attention fusion (SCL-AF).
• Utilized contrastive pretraining with cross-view positives and contralateral hard negatives.
• Implemented lesion-guided tokenization and geometry-biased, bidirectional attention fusion.
• Employed supervised fine-tuning with class-imbalance-aware objectives and regularizers for view consistency and contralateral symmetry.

Main Results:

• SCL-AF achieved ROC-AUC of 0.942, PR-AUC of 0.692, and SEN of 0.631 on the CBIS-DDSM dataset, outperforming baseline models.
• Significant performance gains were observed in the high-specificity regime, particularly for calcification-dominant breasts.
• Ablation studies confirmed the importance of cross-view positives, contralateral negatives, lesion-guided tokens, and bidirectional attention.

Conclusions:

• Encoding mammographic anatomy directly into representation learning and fusion significantly improves AI performance.
• The SCL-AF framework demonstrates potential for enhancing screening mammography triage.
• The model's ability to handle view-invariance and contralateral asymmetry is key to its success.