Adapted foundation models for breast MRI triaging in contrast-enhanced and non-contrast-enhanced protocols

Area of Science:

• Artificial Intelligence in Medical Imaging
• Machine Learning for Radiology
• Breast MRI Analysis

Background:

• Abbreviated breast MRI protocols are increasingly used for screening.
• Accurate triaging of these exams is crucial for efficient workflow.
• AI models offer potential for automated analysis and triage.

Purpose of the Study:

• To evaluate a DINOv2-based medical slice transformer (MST) for triaging abbreviated breast MRI.
• To assess the MST's ability to rule out examinations with Breast Imaging Reporting and Data System (BI-RADS) category ≥4 findings.
• To evaluate performance across contrast-enhanced and non-contrast-enhanced MRI protocols.

Main Methods:

• Retrospective study of 1847 in-house and 924 external abbreviated breast MRI examinations.
• Four abbreviated MRI protocols were tested: T1-weighted early subtraction (T1_sub), diffusion-weighted imaging (DWI₁₅₀₀), DWI₁₅₀₀+T2-weighted (T2w), and T1_sub+T2w.
• Performance assessed using five-fold cross-validation and area under the receiver operating characteristic curve (AUC) at 90%, 95%, and 97.5% sensitivity.

Main Results:

• The T1_sub+T2w protocol achieved an AUC of 0.77±0.04, with no significant differences across protocols.
• At 97.5% sensitivity, T1_sub+T2w had the highest specificity (19%±7%), followed by DWI₁₅₀₀+T2w (17%±11%).
• External validation of T1_sub yielded an AUC of 0.77; 88% of attention maps were rated good or moderate.

Conclusions:

• The MST framework achieved 19% specificity for contrast-enhanced and 17% for non-contrast-enhanced MRI at 97.5% sensitivity.
• This foundation-model-based AI shows potential to support efficient triaging of abbreviated breast MRI.
• The approach may reduce radiologist workload by prioritizing higher-risk studies.