Adapted DINOv2 transformer safely rules out actionable breast MRI findings, yielding 19% specificity at high sensitivity

In 1847 breast MRI exams, a DINOv2-based medical slice transformer triaged cases without suspicious findings (BI-RADS ≥4) at 97.5% sensitivity. Specificity was 19% for contrast-enhanced (T1sub+T2w) and 17% for non-contrast-enhanced (DWI1500+T2w) MRI. External validation AUC 0.77.

• At 97.5% sensitivity, specificity was 19% ± 7% for T1-subtraction plus T2-weighted and 17% ± 11% for diffusion-weighted plus T2-weighted protocols; AUCs did not differ significantly.
• Missed lesions were predominantly <10 mm non-mass enhancements.
• External validation of the T1-subtraction protocol yielded an AUC of 0.77, with 88% of attention maps rated good or moderate.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.