Diagnostic accuracy and translational readiness of deep learning–assisted breast MRI: Systematic review and patient-level HSROC meta-analysis

Abstract

Background: Deep learning (DL) has been proposed to support breast MRI interpretation, but patient-level diagnostic accuracy and translational readiness remain uncertain. J o u r n a l P r e -p r o o fMethods: We searched MEDLINE (Ovid), Embase (Ovid), and Web of Science (January 2010-June 26, 2025) for prospective or retrospective cross-sectional/cohort diagnostic accuracy studies evaluating DL-assisted breast MRI for cancer detection.Eligible studies reported patientlevel outcomes with sufficient information to reconstruct 22 tables; lesion-wise analyses were excluded.Risk of bias and applicability were assessed using QUADAS-2.Translational readiness was summarised descriptively (external validation, interpretability approaches, model availability, workflow integration, and regulatory status).Results: Nine studies (2019-2025; total N=4,472) met eligibility criteria.Sensitivity ranged from 0.88 to 1.00 and specificity from 0.12 to 0.92.HSROC meta-analysis yielded pooled sensitivity 0.94 (95% CI 0.91-0.97)and pooled specificity 0.70 (95% CI 0.49-0.85),with LR+ 3.10 and LR-0.08 (95% CI 0.06-0.11).No study was low risk of bias across all QUADAS-2 domains; 7/9 were high risk in 1 domain.Translational reporting was limited: 3/9 reported external validation, 4/9 reported interpretability methods, and none reported regulatory approval or patient-facing safety/workflow evaluations.Conclusion: DL-assisted breast MRI shows high pooled sensitivity with moderate specificity at the patient level, but methodological weaknesses and limited validation and transparency constrain readiness for practice.Prospective multicentre adjunctive evaluations and reporting aligned with STARD-AI/CLAIM are needed.