Deep Learning in Cardiac Imaging

Abstract

Cardiac imaging -- echocardiography, cardiac MRI, coronary CT angiography, and nuclear perfusion imaging - generates over 120 million studies annually worldwide, each requiring expert interpretation to diagnose conditions ranging from valvular disease to cardiomyopathy to coronary artery stenosis. Deep learning models have demonstrated cardiologist-level performance on specific tasks: left ventricular segmentation, ejection fraction estimation, wall motion abnormality detection, and coronary plaque characterisation. Yet clinical adoption remains limited -- fewer than 5% of cardiology departments use AI-assisted interpretation in routine workflow as of 2024. The barriers are not primarily accuracy but rather integration, generalisability, explainability, and regulatory approval. We present the Cardiac Imaging AI Assessment Framework (CIAAF), evaluating five deep learning architectures -- convolutional neural networks (CNNs), vision transformers (ViTs), U-Net segmentation networks, spatiotemporal recurrent models, and multimodal fusion networks -- across four cardiac imaging modalities (echocardiography, cardiac MRI, coronary CTA, and SPECT myocardial perfusion). Our Cardiac AI Performance Score (CAPS) measures diagnostic accuracy, generalisability across institutions, inference speed, explainability, and clinical workflow integration. Vision transformers achieve the highest CAPS (0.922) through superior cross-institutional generalisability (AUC degradation < 2% across five external sites) and attention-based explainability, while U-Net architectures achieve the highest segmentation accuracy (Dice 0.962) for ventricular volumetry.