Uncertainty-Aware Skull Reconstruction

Project page

Uncertainty-Aware Skull Reconstruction

Uncertainty estimation and probabilistic skull shape reconstruction using Bayesian neural networks.

Jianning Li, Agniva Sengupta, and Stefan Zachow

Scientific Reports, Volume 16, Article 16383, 2026

Paper Code Overview BibTeX

Probabilistic 3D reconstruction examples and uncertainty-aware outputs from the project.

Overview

Probabilistic reconstruction for missing skull anatomy.

This work studies uncertainty-aware deep learning for 3D skull reconstruction in cranioplasty. The method uses a 3D Bayesian U-Net to jointly predict missing skull structures and estimate epistemic uncertainty through Monte Carlo approximation with Flipout layers. By pairing voxel-grid reconstruction with uncertainty maps, the system provides both a candidate skull shape and spatial evidence about where the model is more or less confident.

Method

Bayesian neural networks for 3D shape completion.

The codebase supports cranial completion, facial bone completion, and skull shape super-resolution. Training and inference operate on volumetric NIfTI data, while visualization scripts map voxel predictions and uncertainty values back to inspectable 3D surfaces.

3D Bayesian U-Net

Flipout layers model weight uncertainty while preserving the encoder-decoder structure used for dense volumetric prediction.

Monte Carlo inference

Repeated stochastic forward passes produce mean reconstructions and epistemic variance maps.

Surface visualization

Prediction and uncertainty volumes can be converted to meshes for qualitative inspection and analysis.

Results

Qualitative reconstruction examples.

Examples across cranial defects, facial defects, and super-resolution scenarios.

Cranial defect reconstruction example 1 — **Cranial defect reconstruction**Bayesian reconstruction of missing cranial anatomy with uncertainty-aware output.

Cranial defect reconstruction example 2 — **Cranial defect reconstruction**Second skull completion example demonstrating volumetric shape recovery.

Skull shape super-resolution example 1 — **Skull shape super-resolution**Super-resolution output for skull anatomy.

Skull shape super-resolution example 2 — **Skull shape super-resolution**Additional super-resolution example from the repository materials.

Facial defect reconstruction example — **Facial defect reconstruction**Example output for facial bone completion.

Resources

Implementation and reproducibility.

The official implementation includes model definitions, data loaders, training and inference scripts, evaluation utilities, and visualization tools for mesh-based inspection of uncertainty.

Open Repository Read Paper

Citation

Cite this work.

@article{li2026uncertainty,
  title = {Uncertainty estimation and probabilistic skull shape reconstruction using Bayesian neural networks},
  author = {Li, Jianning and Sengupta, Agniva and Zachow, Stefan},
  journal = {Scientific Reports},
  volume = {16},
  number = {16383},
  year = {2026},
  month = {May},
  doi = {10.1038/s41598-026-54679-7},
  url = {https://www.nature.com/articles/s41598-026-54679-7}
}