Reproduce “Unsupervised Multi-Modal Medical Image Registration via Discriminator-Free Image-to-Image Translation” 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
Zekang Chen, Jia Wei, Rui Li
An unsupervised multi-modal medical image registration framework employs a discriminator-free image-to-image translation network, achieving state-of-the-art registration accuracy with an average Dice Similarity Coefficient of 0.772 on the CT-MR thorax-abdomen dataset, by explicitly preserving anatomical shapes and avoiding artifacts.
Key insights
- Discriminators in GAN-based image translation, while promoting visual realism, can inadvertently introduce shape inconsistencies and artifacts that are detrimental to medical image registration accuracy.
- Removing the discriminator and leveraging contrastive learning (PatchNCE) can effectively preserve anatomical shapes during modality translation, which is crucial for registration tasks.
- Explicitly combining local alignment (for fine-grained details) and global alignment (for overall structural correspondence) losses guides the registration network to achieve comprehensive and accurate deformations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.