DeepSDF and MCMC Enable Reliable Cardiac Shape Reconstruction
A probabilistic framework quantifies uncertainty in heart shape reconstruction from sparse point clouds.
Cardiac shape reconstruction from sparse or noisy data (e.g., point clouds) is often unreliable due to prior-driven methods that ignore uncertainty. Researchers Jan Verhülsdonk and colleagues propose a probabilistic framework that combines Deep Signed Distance Functions (DeepSDF) with Markov Chain Monte Carlo (MCMC) sampling. DeepSDF implicitly represents cardiac geometry as the zero-level set of a neural network conditioned on learned latent codes. The model handles multi-surface reconstruction of the left and right ventricles. By interpreting the reconstruction loss as a log-likelihood, they perform Bayesian inference in the latent space, generating both maximum a posteriori (MAP) and full posterior-sampled reconstructions.
Experiments on a public cardiac dataset show that the framework produces accurate segmentations and well-calibrated uncertainty estimates. This is a critical step toward reliable clinical use, where knowing the confidence of a reconstruction can guide treatment decisions. The method is atlas-based but adds probabilistic rigor, making it suitable for downstream tasks like surgical planning, risk stratification, and longitudinal monitoring. By quantifying uncertainty, clinicians can trust reconstructions even from sparse or noisy input, potentially reducing the need for repeated scans or invasive measurements.
- Combines Deep Signed Distance Functions (DeepSDF) with MCMC sampling for probabilistic cardiac shape reconstruction.
- Performs Bayesian inference in latent space, yielding both MAP and posterior-sampled reconstructions with uncertainty estimates.
- Achieves accurate multi-surface reconstruction of left and right ventricles on a public cardiac dataset with well-calibrated confidence.
Why It Matters
Clinicians can now assess reconstruction reliability from sparse data, improving diagnosis and surgical planning.