To leverage an AI vision language model for interpreting cardiac magnetic resonance imaging (MRI) scans to improve efficiency and quality of reporting.
Approach:
Study Design: The study utilized CMR images and cardiologists' text reports from Cleveland Clinic and University Hospital of Dijon, trained on over 11,028 deidentified patient studies.
Model Development: The CMR contrastive language image pretraining (CMR-CLIP) model was developed to treat MRI scans as video-like sequences, learning from over a million images.
Evaluation: The model was evaluated on tasks including classification of cardiomyopathies, prediction of ejection fraction, and CMR report generation.
Key Findings:
CMR-CLIP achieved accuracies of 88.5% for nonischemic cardiomyopathy, 88.0% for ischemic cardiomyopathy, 96.2% for cardiac amyloidosis, and 98.6% for hypertrophic cardiomyopathy, as reported by the study.
The model identified cardiac conditions in a zero-shot setting without needing expert annotations, according to the research findings.
It can search large databases using natural language, aiding in quick comparisons of rare or complex conditions, as highlighted in the study.
Interpretation:
Limitations:
The model was tested on retrospective cases and needs validation for clinical reporting, as noted in the study.
Continual refinement of the tool is necessary for optimal performance, according to the researchers.
Conclusion:
The CMR-CLIP model could enhance the efficiency and quality of cardiac MRI reporting, serving as a teaching tool and supporting automated clinical decision-making, as suggested by the study.
