Reproduce “Adaptive Conditional Contrast-Agnostic Deformable Image Registration with Uncertainty Estimation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yinsong Wang, Xinzhe Luo, Siyi Du, Chen Qin
Imperial College London researchers developed AC-CAR, an Adaptive Conditional Contrast-Agnostic Deformable Image Registration framework that aligns medical images with arbitrary contrasts and provides uncertainty estimations. The framework demonstrated improved Dice scores (e.g., 0.808 on CamCAN T1w →T2w) and maintained high accuracy on unseen contrasts, while its uncertainty maps accurately correlated with misregistration.
Key insights
- Learning robust contrast-invariant features can be effectively achieved by adaptively modulating feature representations based on low-frequency image content and regularizing latent space features across augmented contrasts.
- Uncertainty estimation for deformable registration can be made contrast-agnostic by sharing the encoder weights with the registration network, leveraging its contrast-invariant feature learning.
- Simplifying supervision using a mono-contrast similarity loss on pre-augmented mono-contrast images, combined with effective contrast augmentation, enables generalizable multi-contrast registration.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.