Clinical Scorecard: Anatomy-Informed Visual Prompt Adjustment for Enhanced Cross-Modal Understanding of Breast Cancer

At a Glance

Key Highlights

• A-VPT framework incorporates glandular, fatty, and ductal anatomical embeddings into frozen Vision Transformer models for tissue-aware prompt tuning.
• Cross-modal contrastive alignment harmonizes anatomical semantics across mammography, ultrasound, and MRI, enhancing multi-domain generalization.
• Achieves state-of-the-art lesion classification and segmentation performance using less than 2% of parameters compared to full fine-tuning, with interpretable attention aligned to radiological structures.

Guideline-Based Recommendations

Diagnosis

• Utilize multi-modal imaging (mammography, ultrasound, MRI) for comprehensive breast cancer lesion assessment.
• Incorporate anatomical priors to improve cross-domain consistency and lesion detection accuracy.

Management

• Adopt advanced AI frameworks like A-VPT that embed anatomical knowledge to enhance imaging interpretation efficiency and reliability.

Monitoring & Follow-up

• Leverage interpretable AI attention maps consistent with anatomical structures to support ongoing lesion evaluation.

Risks

• Be aware of potential limitations in data-driven models lacking anatomical context leading to reduced generalizability across imaging modalities.

Patient & Prescribing Data

Breast cancer patients undergoing imaging diagnostics

Embedding anatomical priors into AI models can improve diagnostic accuracy and interpretability, potentially guiding personalized treatment planning.

Clinical Best Practices

• Incorporate domain-specific anatomical information into AI model training for breast cancer imaging to enhance diagnostic performance.
• Use parameter-efficient fine-tuning methods to adapt large pre-trained models with minimal computational resources.
• Apply cross-modal alignment strategies to ensure consistent interpretation across different imaging modalities.

References

• Sung et al. Global cancer statistics 2020
• Moreira et al. INbreast dataset
• Al-Dhabyani et al. BUSI dataset
• Saha et al. Duke-Breast-MRI dataset