| Keywords [eng] |
Artificial Intelligence, Cardiovascular Magnetic Resonance Imaging, Wall Thickness Measurement, Cardiac Volumetry, Myocardial Mass, Myocardial Scar Detection, Deep Learning, Segmentation, Clinical Workflow Optimization, Observer Variability. |
| Abstract [eng] |
Objectives: Currently the analysis of cardiovascular magnetic resonance (CMR) images is entirely dependent on the expertise of individual trained professionals and thus is time consuming and introduces observer variation. Artificial intelligence (AI) is becoming more and more relevant when it comes to addressing these limitations, allowing for automation of repetitive processes such as analysis of cardiac morphology, function and myocardial structure. The aim of this review was to assess the role of AI in CMR imaging, with a specific focus on three domains: volumetry and mass, myocardial scar detection and quantification, and myocardial wall thickness measurements. To accomplish this task a systematic review was conducted using the PRISMA technique including high-quality studies published between November 2020 and January 2024. Results: In total, 15 studies were analyzed. Compared to human expert evaluation, CMR-based AI systems have been shown to significantly reduce intra- and interobserver variabilities, with segmentation Dice scores of ≥ 0.90 for both left and right ventricular volumetry and ICCs up to 0.94 for left ventricular volume measurements. For myocardial wall thickness, AI demonstrated improved test–retest reproducibility with a CoV of 4.3%, compared to 5.7–12.1% for human observers. In the field of tissue characterization, AI solutions allow assessment of myocardial fibrotic changes without contrast media; for instance, Virtual Native Enhancement (VNE) achieved an ICC of 0.94 compared to LGE for scar burden. These advances are valuable in clinical practice for decision-making and reduce required sample sizes in research. AI-based analysis also significantly improves efficiency, reducing analysis time from 10–15 minutes to under 30 seconds in some implementations. Despite advancement of AI based CMR analysis the need for expert oversight and manual review in more complex cases is still needed, as well as generalizability, data standardization, and ethical issues still limits widespread use of AI solutions in CMR image analysis. Conclusions: Across all three analyzed domains, AI consistently demonstrated improvements in reproducibility, efficiency, and diagnostic consistency when compared to traditional manual human expertise-based methods. Successful integration of AI into clinical CMR workflows will depend in the future on further validation in larger patient cohorts, cross-center generalizability, and sustained expert supervision. |
