Towards Robust General Medical Image Segmentation

MICCAI 2021

logo CinfonIA

Fork me on GitHub

Adversarial Robustness in Neural Networks

Neural networks have demonstrated outstanding performance in recognition tasks. However, these methods are highly vulnerable to imperceptible corruptions in the input images.


Ground truth
gif gt

Prediction on a clean image
gif pred
Prediction on adversarial example
gif pred-adv

RObust General Segmentation (ROG)

We introduce a lattice architecture for general medical image segmentation

ROG

Adversarial Robusntess Benchmark

We introduce a new benchmark to reliably assess adversarial robustness on the Medical Segmentation Decathlon. Our benchmarck includes 5 state-of-the-art adversarial attacks that we expanded to the domain of 3D segmentation:

  • PGD
  • APGD-CE
  • APGD-DLR (only for non-binary tasks)
  • FAB
  • Square Attack
Benchmark graph

Paper

You can get access to our full paper here. If you find our work useful, please use the following BibTeX entry for citation:

@inproceedings{daza2021towards,
  title={Towards Robust General Medical Image Segmentation},
  author={Daza, Laura and P{\'e}rez, Juan C and Arbel{\'a}ez, Pablo},
  booktitle={MICCAI},
  year={2021}
}

Team

Photo Laura Laura Daza
photo Juan C Juan C. Pérez
Photo Pablo Pablo Arbeláez