Project Report
Few-Shot 2D Echo to 3D Cardiac Reconstruction via Neural Implicit Priors
This project reconstructs 3D left-ventricular geometry from sparse 2D echocardiographic supervision using an implicit neural representation. The strongest full-volume overlap in the saved summaries is the mixed run without stratifiers, with 0.8643 Dice and 0.7658 IoU.
Key results
| Setting | 2D Dice | 2D IoU | 3D Dice | 3D IoU |
|---|---|---|---|---|
| Mixed stratified ED-healthy | 0.9458 | 0.9007 | 0.8491 | 0.7422 |
| Meta / Reptile, after refinement | 0.9540 | 0.9143 | 0.8638 | 0.7649 |
| Mixed, no stratifiers | 0.9505 | 0.9085 | 0.8643 | 0.7658 |
What was built
- Coordinate-based implicit neural representation for 3D occupancy.
- Learnable view-specific pose parameters for each 2D slice.
- Mixed and Meta initialization paradigms for few-shot adaptation.
- Interactive Plotly viewers for comparing reconstructed volumes against segmentation.
The codebase uses the repo-local dataset at cap-mitea/mitea.