A next-generation artificial intelligence system can analyze complex heart scans better than existing models without the need for laboriously, manually labeled training data.
The system, outlined in Nature Communications, sets the scene for multimodal learning approaches to be further integrated into medical imaging, with the potential to improve diagnoses and patient outcomes.
The vision-language self-supervised learning framework for cardiac magnetic resonance (CMR) imaging uses contrastive language image pretraining (CLIP) and treats scans as videos of the beating heart.
The novel CMR-CLIP system outperformed existing models by 35% and was better at identifying common pathologies such as myocardial fibrosis and left ventricular hypertrophy.
It also beat other models at common computational tasks such as the retrieval of CMR studies or radiology reports and downstream disease classification tasks.
“Systems like CMR-CLIP have the potential to support clinicians through automated screening, and interpretation support, particularly in settings where expert readers are limited,” explained researcher David Chen, PhD, of Cleveland Clinic.
“Such reader assistant tools are critical to improving patient access to this powerful diagnostic technology.”
Cardiac magnetic resonance imaging is the definitive way to diagnose several cardiac diseases including valvular pathologies, cardiomyopathies, pericardial and aortic diseases.
However, interpreting and documenting each exam takes a great deal of time due to the amount of information collected in each CMR exam—often more than 40 minutes per study.
Vision-language models trained using self-supervised learning are therefore crucial to reduce dependency on large volumes of labeled data.
However, conventional self-supervised approaches that rely on precise image-text pairing are not always feasible for CMR, given that it is able to visualize cardiac anatomy, physiology, and microstructure in a single exam.
Unlike either generalists and other biomedical domain-specific models, which are trained using individual images or limited views, CMR-CLIP incorporates a wide variety of standard cardiac views and image types that represent of morphology, function, and myocardial viability.
The vision language model connects images and associated reports, treating the various views of the heart and image types as a sequence of images in video format.
The model was trained on over a million images from over 10,000 unique studies at a single institution and performed well on evaluation in both on internal and external datasets.
The researchers said it achieved “remarkable performance” at real-world clinical tasks, reaching accuracies of 88.5% for non-ischemic cardiomyopathy, 88.0% for ischemic cardiomyopathy, 96.2% for cardiac amyloidosis, and 98.6% for hypertrophic cardiomyopathy.
“This work demonstrates that domain-specific foundation models can significantly outperform general-purpose AI systems in specialized clinical applications,” said researcher Ding Zhao, PhD, Carnegie Mellon University.
“By designing models that reflect the structure and complexity of cardiac MRI data, rather than adapting generic image models, we can unlock new levels of performance and clinical utility.”
The post Next-Generation Cardiac AI System Outperforms Existing Models appeared first on Inside Precision Medicine.