Echo4DIR: 4D Heart Reconstruction from 2D Echo Videos
A new AI framework, Echo4DIR, reconstructs 4D (3D+t) cardiac geometry from sparse 2D echocardiography videos. The method addresses geometric ambiguity and temporal discontinuity by learning 3D shape priors from statistical shape models via a cardiac conditional SDF. An Epipolar Mask Encoder module with epipolar cross attention fuses multi-view features. A self-supervised SDF-tailored differentiable rendering strategy adapts to patient-specific shapes using uncalibrated clinical masks without 3D ground truth. The implicit representation ensures anatomically reliable geometry at arbitrary resolutions. The framework is detailed in arXiv:2605.22066.
Key facts
- Echo4DIR reconstructs 4D cardiac geometry from 2D echocardiography
- Uses statistical shape models for 3D shape priors
- Employs Epipolar Mask Encoder with epipolar cross attention
- Self-supervised differentiable rendering adapts to patient-specific shapes
- No 3D ground truth required for adaptation
- Implicit representation overcomes sparse observations
- Published on arXiv with ID 2605.22066
- Addresses geometric ambiguity and temporal discontinuity
Entities
Institutions
- arXiv