Clinical Report: Multi-Modal Network Utilizing Cross-Attention for Breast Ultrasound
Overview
The Cross-Attention Guided Network (CGA-Net) demonstrates improved diagnostic accuracy in breast ultrasound by integrating visual data with clinical descriptors. It achieved an Out-Of-Fold AUC of 0.905 and outperformed traditional models, suggesting a significant reduction in false positives.
Background
Breast cancer is a leading cause of cancer-related mortality among women, making accurate diagnosis crucial for improving survival rates. Breast ultrasound is a primary screening tool, yet its effectiveness is often limited by operator subjectivity and inter-observer variability. The integration of AI and structured clinical information may enhance diagnostic accuracy and reduce unnecessary procedures.
Data Highlights
Model
Out-Of-Fold AUC
Overall Accuracy
Specificity
Sensitivity
CGA-Net (trained from scratch)
0.905
0.857
0.831
0.898
CGA-Net (pre-trained)
0.915
N/A
N/A
N/A
Clinical-only baseline
0.890
N/A
N/A
N/A
Image-only baseline
0.795
N/A
N/A
N/A
Key Findings
The CGA-Net model integrates visual data with clinical descriptors to enhance diagnostic accuracy.
It achieved an Out-Of-Fold AUC of 0.905 and an overall accuracy of 0.857.
Specificity was recorded at 0.831, while sensitivity reached 0.898.
A pre-trained version of CGA-Net outperformed the clinical-only and image-only baselines.
Attention map visualizations indicated alignment with expert focus on tumor periphery.
Clinical Implications
CGA-Net provides a robust tool for clinicians, potentially reducing diagnostic variability in breast ultrasound. Its ability to integrate structured clinical information may enhance decision-making and improve patient outcomes.
Conclusion
CGA-Net represents a significant advancement in breast ultrasound diagnostics, effectively combining visual and clinical data to improve accuracy and reduce false positives.
Akeso’s PD-1/VEGF bispecific beats PD-1 plus chemo in lung cancer, Amgen wins EC approval for its DLL3-targeted T-cell engager, and Aragen scales Renaissance’s Fast Track anti-GD2 antibody
