Reproduce “Echo4DIR: 4D Implicit Heart Reconstruction from 2D Echocardiography Videos” on OpenResearch

About this paper

Yanan Liu, Qinya Li, Hao Zhang, Kangjian He, Xuan Yang, Hao Li, Dan Xu, Lei Li

Echo4DIR reconstructs 4D cardiac geometry from sparse 2D echocardiography videos using conditional Signed Distance Functions and a test-time optimization framework, reducing 3D geometric error by over 50% compared to previous methods on a synthetic dataset. The framework achieves 98.35% Dice coefficient for 2D overlap with clinical segmentations and maintains topological consistency through a novel Radial SDF Alignment strategy.

Key insights

• Leveraging implicit neural representations (SDFs) with learned shape priors from statistical models can effectively infer complete 3D cardiac geometry from sparse 2D views, addressing geometric ambiguity.
• A differentiable rendering approach allows for robust self-supervised adaptation of implicit 3D shapes to real 2D clinical data during test-time, overcoming the synthetic-to-real domain gap without requiring 3D ground truth.
• Explicitly aligning propagated vertices to the current frame's implicit surface via Radial SDF Alignment effectively maintains topological consistency and smooth motion across time, mitigating temporal discontinuity artifacts.

Summary by alphaXiv.