To enhance cross-modal understanding and detection of breast cancer by integrating anatomical priors into visual prompt tuning.
Key Findings:
A-VPT achieves state-of-the-art performance in lesion classification and segmentation.
Utilizes less than 2% of the tunable parameters required for full fine-tuning.
Anatomy-guided prompts provide interpretable attention patterns consistent with radiological structures.
Interpretation:
Embedding anatomical priors into prompt tuning enhances model efficiency, generalization, and interpretability in breast cancer detection across imaging modalities.
Limitations:
The code supporting the findings is not publicly available at the moment.
Further validation on diverse datasets may be needed to generalize findings.
Conclusion:
A-VPT represents a significant advancement in integrating anatomical knowledge into deep learning for breast cancer imaging, improving both performance and interpretability.
