Cross-attention guided multi-modal network for breast ultrasound diagnosis incorporating objective clinical semantics
-
By
-
Qin Sun
-
Xiaoman Wu
-
Ju Chen
-
Yuhang Zhang
-
Chao Zhou
-
Zheng Zhu
-
June 9, 2026
-
Clinical Scorecard: Multi-Modal Network Utilizing Cross-Attention for Objective Clinical Semantics in Breast Ultrasound Diagnosis
At a Glance
| Category | Detail |
|---|
| Condition | |
| Key Mechanisms | Cross-Attention Guided Network (CGA-Net) integrates visual data with clinical descriptors. |
| Target Population | |
| Care Setting | |
Key Highlights
- CGA-Net achieved an Out-Of-Fold AUC of 0.905 and overall accuracy of 0.857.
- The model demonstrated high sensitivity (0.898) and specificity (0.831).
- CGA-Net outperformed both clinical-only and image-only baselines.
Guideline-Based Recommendations
Diagnosis
- Utilize structured clinical descriptors alongside imaging data.
Management
Monitoring & Follow-up
Risks
- Potential for false positives if not properly calibrated.
Patient & Prescribing Data
252 patients (98 malignant, 154 benign) from the BrEaST dataset.
CGA-Net provides a data-efficient solution for distinguishing between benign and malignant lesions.
Clinical Best Practices
Related Resources & Content
