Eur Radiol. 2026 Aug 13. doi: 10.1007/s00330-026-12782-3. Online ahead of print.
ABSTRACT
OBJECTIVES: To evaluate a DINOv2-based medical slice transformer (MST) for triaging abbreviated breast MRI by ruling out examinations with suspicious findings that are immediately actionable (Breast Imaging Reporting and Data System [BI-RADS] ≥ 4) across contrast-enhanced and non-contrast-enhanced protocols.
MATERIALS AND METHODS: This institutional review board-approved retrospective study included 1847 single-breast MRI examinations (377 BI-RADS ≥ 4) from an in-house dataset and 924 from an external dataset (Duke). Four abbreviated protocols were tested: T1-weighted early subtraction (T1sub), diffusion-weighted imaging with b = 1500 s/mm² (DWI1500), DWI1500 + T2-weighted (T2w), and T1sub + T2w. Performance was assessed at 90%, 95%, and 97.5% sensitivity using five-fold cross-validation and area under the receiver operating characteristic curve (AUC). AUC differences were compared with the DeLong test. False negatives were characterized, and attention maps were rated in the external dataset.
RESULTS: A total of 1448 female patients (mean age, 49 ± 12 years) were included. T1sub + T2w achieved an AUC of 0.77 ± 0.04; DWI1500 + T2w, 0.74 ± 0.04, with no significant differences across protocols. At 97.5% sensitivity, T1sub + T2w had the highest specificity (19% ± 7%), followed by DWI1500 + T2w (17% ± 11%). At 95% and 97.5% sensitivity, missed lesions were predominantly < 10 mm, mainly non-mass enhancements. External validation of the T1sub protocol yielded an AUC of 0.77, with 88% of attention maps rated good or moderate.
CONCLUSION: At 97.5% sensitivity, the MST framework triaged cases without BI-RADS ≥ 4, achieving 19% specificity for contrast-enhanced and 17% for non-contrast-enhanced MRI. These findings highlight the potential of foundation-model-based AI to support efficient triaging of abbreviated breast MRI protocols.
KEY POINTS: Question Can an adapted DINOv2-based MST accurately stratify abbreviated breast MRI examinations to support triaging across contrast-enhanced and non-contrast-enhanced imaging protocols? Findings The model achieved 19% specificity for contrast-enhanced and 17% for non-contrast-enhanced MRI at 97.5% sensitivity; external validation of the T1sub protocol yielded an AUC of 0.77. Clinical relevance A foundation-model-based triage approach may reduce radiologist workload in breast MRI screening by identifying examinations unlikely to contain suspicious findings, including on non-contrast-enhanced protocols, thereby enabling prioritization of higher-risk studies.
PMID:42593493 | DOI:10.1007/s00330-026-12782-3
