Laplacian Segmentation Networks Improve Epistemic Uncertainty Quantification

Kilian Zepf*, Selma Wanna*, Marco Miani, Juston Moore, Jes Frellsen, Søren Hauberg, Frederik Warburg, Aasa Feragen (MICCAI 2024)

$^*$ denotes equal contribution

[Paper on Arxiv]

This repository contains an implementation of the proposed model class as well as the benchmarks presented in the paper. The code is based on PyTorch.

$\qquad$

Figure: Model overview Laplacian Segmentation Network - uncertainty measures are calculated by approximating expectations by Monte Carlo-sampling mean networks from the Laplace approximation $q(θ_*)$ and predicting the respective logit distributions $p(η|x,θ)$ for $x$.

Citation

If our method is helpful for your own research, please consider citing our MICCAI 2024 paper:

@inproceedings{zepf2024laplacian,
  title={Laplacian Segmentation Networks Improve Epistemic Uncertainty Quantification},
  author={Zepf, Kilian and Wanna, Selma and Miani, Marco and Moore, Juston and Frellsen, Jes and Hauberg, S{\o}ren and Warburg, Frederik and Feragen, Aasa},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={349--359},
  year={2024},
  organization={Springer Nature Switzerland Cham}
}