Cross-Center Vision-Language Transformer for Robust Mammography-Based Breast Cancer Diagnosis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Deep learning for mammography shows promise but struggles with performance variability across clinical settings.
• Reliance on visual data alone limits robustness due to diverse imaging protocols, scanners, and patient populations.

Purpose of the Study:

• To develop a robust deep learning framework for breast cancer diagnosis that overcomes limitations of visual-only approaches.
• To enhance diagnostic accuracy and reliability across diverse clinical environments by integrating imaging and text data.

Main Methods:

• Introduced the Cross-Center Vision-Language Transformer (CC-VLT) framework.
• Integrated a vision transformer for mammograms and a text transformer for clinical descriptors.
• Employed bi-directional cross-modal attention and cross-center feature regularization to address domain shifts.

Main Results:

• CC-VLT achieved 90.7% accuracy and high ROC-AUC scores (0.951 intra-center, 0.912-0.934 cross-center) on multiple datasets.
• Demonstrated superior performance compared to baseline models in both intra- and cross-center evaluations.
• Improved reliability of malignancy probability predictions, indicated by reduced Expected Calibration Error and Brier Score.

Conclusions:

• The CC-VLT framework offers a robust solution for breast cancer diagnosis in diverse clinical settings.
• Integrated vision-language models with cross-center regularization establish a new benchmark for reliable mammography analysis.
• This approach enhances diagnostic performance and reliability, addressing key challenges in real-world clinical deployment.