Deep learning for cardiac MRI: performance evidence and barriers to clinical integration. A Systematic Review and Meta-Analysis

Abstract Aims This systematic review and meta-analysis aimed to evaluate the current evidence on the use of deep learning in cardiac magnetic resonance imaging, focusing on image segmentation, prediction, and diagnosis. Methods and results A systematic search of Medline, Web of Science, Embase, and Scopus identified studies published between 2020 and 2025. Eligible studies comprised deep learning-based segmentation, prediction, or diagnosis of cardiac magnetic resonance images. MetaDisc version 1.4 was used for statistical analysis, with a P < 0.05 and an I2 ≥ 75% used as the thresholds for statistical significance and high heterogeneity, respectively. From 1510 retrieved articles, 62 studies met the inclusion criteria, and 12 studies were included in the meta-analysis. Most studies targeted segmentation (n & 45), with fewer addressing diagnosis (n & 9), and prediction (n & 28). Supervised learning predominated (91.94%), and U-Net was the most common architecture (70.97%). Mean Dice score (15 studies) was 0.91 ± 0.03, whereas mean Hausdorff distance (six studies) was 8.99 ± 6.45 mm. Diagnosis and prediction achieved pooled sensitivity of 0.94 (95% CI: 0.92–0.96), specificity of 0.91 (95% CI: 0.89–0.93), and AUC of 0.9831, indicating excellent discriminative ability. Segmentation models reached pooled sensitivity of 1.00 (95% CI: 0.99–1.00) and specificity of 0.98 (95% CI: 0.98–0.99). The AUC from the SROC analysis was 0.9940, confirming exceptional segmentation accuracy. Conclusion Deep learning models show excellent performance in cardiac magnetic resonance segmentation and diagnosis, often matching or exceeding manual analysis, indicating strong potential for clinical adoption. This systematic review was registered in the International Prospective Register of Systematic Reviews (PROSPERO) with the registration number CRD42023439659.