Master Thesis: 4D JEPA (Joint Embedding Predictive Architectures) for Cardiac MRI

Summary

Joint Embedding Predictive Architectures (JEPA) predict latent representations instead of reconstructing pixels. This thesis adapts JEPA to cine cardiac MRI in its native 4D form (3D space plus time) and compares 2D, 2D+T, 3D and 4D inputs on segmentation and functional tasks to find which format justifies its computational cost. Submit your application at the TUM Portal.