Influence of virtual dynamic contrast enhanced breast MRI on lesion size, BI-RADS scores and reader confidence

Abstract

Abstract Background Breast MRI routinely uses gadolinium-based contrast agents (GBCAs) for dynamic contrast-enhanced (DCE) imaging. However, GBCA administration entails periprocedural effort and environmental impact, prompting exploration of contrast-agent-free approaches. Deep learning-based virtual DCE (vDCE) offers a potential solution. This study evaluated vDCE against multiparametric DCE protocols, focusing on BI-RADS, background parenchymal enhancement (BPE), and lesion size. Materials and methods This IRB-approved retrospective study included 592 3 T examinations divided into training, validation, and test sets. A 2D U-Net generated vDCE series from unenhanced T1-weighted, T2-weighted, and diffusion-weighted acquisitions. Two board-certified radiologists evaluated vDCE and DCE examinations using a BI-RADS-adapted reading protocol. Diagnostic and ordinal outcomes were compared using McNemar and Wilcoxon signed-rank tests with Bonferroni correction. Results Mean BI-RADS-based diagnostic accuracy was 76% for vDCE and 77% for DCE. Mean sensitivity was 92% vs 96%, and mean specificity was 65% vs 64%, respectively, without significant differences (both readers, p = 1.0). The mean paired lesion-size difference (DCE minus vDCE) was −2.2 mm in the holdout test set and 0.5 mm in the secondary test set. BPE (R1 p = 1.0; R2 p = 1.0) and artefact ratings (R1 p = 1.0; R2 p = 1.0) did not differ significantly. Reader confidence was high for both protocols, although lower for vDCE for one reader ( p = 0.072). Conclusion vDCE showed four-percentage-point lower mean sensitivity and small, cohort-dependent lesion-size biases relative to GBCA-enhanced imaging. Further research is warranted to determine its potential role as a contrast-agent-free approach for selected breast imaging indications. Key Points Question Breast MRI relies on GBCAs, creating procedural and environmental burdens, while clinically validated contrast-agent-free approaches for BI-RADS-based assessment remain unavailable . Findings Virtual DCE MRI showed lower sensitivity than standard DCE, no significant difference in BPE, but more variable reader confidence and lesion size presentation . Clinical relevance Virtual DCE breast MRI currently demonstrates limited diagnostic performance and requires further optimisation, while additional clinical validation studies are needed to determine its potential as a non-contrast alternative .