Few-Shot 2D Echo to 3D Cardiac Reconstruction via Neural Implicit Priors

summary

This course project built a framework for reconstructing patient-specific 3D ventricular geometry from sparse 2D echocardiography. The key idea was to combine a learned global shape prior with test-time refinement so the model could stay data-efficient without losing patient-specific detail.

problem

Echocardiography is widely available, but most routine exams provide only a small number of 2D views. Recovering a high-fidelity 3D cardiac shape from that sparse data is hard because large parts of the anatomy are never directly observed.

approach

The project supports three operating modes to balance speed and accuracy:

• Local: train a fresh INR from scratch for every patient
• Global: learn one shared prior across the training population
• Mixed: start from a global prior and refine it at test time on sparse patient-specific slices

The technical stack combines:

• Representation: coordinate-based MLP representing occupancy / signed-distance geometry
• Differentiable rendering: a vectorized projection layer mapping the 3D implicit shape to clinical slice planes
• Pose refinement: joint optimization of rigid slice pose and shape
• Meta-learning: a Reptile-style initialization that adapts quickly to new cardiac geometries

result

• Test-time refinement significantly improved patient-specific reconstruction compared with static inference
• Population-level shape priors regularized anatomy in unobserved regions
• Automated slice selection improved the diagnostic consistency of the sparse views
• Evaluation used 3D Dice, IoU, and clinically meaningful ventricular volume estimates

The method bridges the gap between widely available 2D ultrasound and expensive 3D imaging. By leveraging test-time optimization, the framework can recover plausible 3D meshes from as few as three standard views while keeping the final shape tied to the observed clinical data.

Collaborators: Vivek Dhara and Vaibhav Parekh

limitations

• Reconstruction quality still depends on slice selection and pose quality
• Sparse-view supervision leaves some ambiguity in unobserved regions
• The framework is a research prototype rather than a clinically validated reconstruction tool

next steps

• compare the hybrid mode more rigorously against newer Gaussian-field variants
• improve robustness to poor acquisition geometry
• evaluate on larger and more heterogeneous echo cohorts