Deep Learning for Cardiac MRI Segmentation and Registration

This technical development study presents a novel deep learning framework combining a U-Net architecture with mesh deformation for automated cardiac MRI segmentation and registration, evaluated on 45 cardiac MRI datasets from a public benchmark.

The proposed combined U-Net plus mesh deformation approach achieved high geometric accuracy (Dice coefficient >0.90) for segmentation of cardiac chambers and myocardium, with improved boundary delineation compared with standard U-Net alone. The mesh deformation component enabled smooth surface extraction aligned with cardiac anatomy, which is relevant for downstream biomechanical modeling and analysis.

Automated cardiac segmentation is a technically important problem: accurate, reproducible quantification of cardiac volumes and function from MRI requires precise delineation of ventricular boundaries that is laborious when performed manually. Deep learning approaches now routinely achieve radiologist-level segmentation accuracy in controlled benchmark evaluations.

The clinical relevance to coronary artery disease assessment is indirect: cardiac MRI measures cardiac function, volumes, and myocardial viability (through late gadolinium enhancement) rather than coronary anatomy directly. Improved automated segmentation reduces the time burden of MRI analysis and enables more consistent quantification of function in research and clinical workflows.

We rate the evidence limited. A technical deep learning development study demonstrating improved cardiac MRI segmentation accuracy — methodologically sound within its domain but peripheral to the clinical evidence base for coronary artery disease diagnosis and treatment.