CaLID framework reconstructs 3D cardiac volumes 24x faster from sparse MRI slices

Cardiac Magnetic Resonance (CMR) imaging relies on sparse 2D short-axis slices to assess heart health, but reconstructing complete 3D volumes from these limited views has been a persistent challenge. Existing methods suffer from rigid interpolation (linear/spherical), high computational costs, and dependence on additional semantic inputs like segmentation masks or motion fields—making them less practical for clinical workflows.

To overcome these bottlenecks, Niklas Bubeck and colleagues from multiple institutions introduce CaLID, a cardiac-specific diffusion model that learns interpolation directly from data. By operating entirely in latent space, CaLID achieves a 24-fold speedup in upsampling time compared to prior approaches while maintaining—and often surpassing—state-of-the-art reconstruction quality. The framework requires only sparse 2D CMR images as input, eliminating the need for auxiliary semantic guidance and simplifying clinical deployment. In volumetric evaluations and downstream segmentation tasks, CaLID consistently outperformed baselines. Furthermore, the team extended CaLID to 2D+T data, enabling spatiotemporal reconstruction with enhanced temporal coherence—a critical feature for dynamic cardiac analysis.