Reproduce “Reverse Imaging for Wide-spectrum Generalization of Cardiac MRI Segmentation” 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

Yidong Zhao, Peter Kellman, Hui Xue, Tongyun Yang, Yi Zhang, Yuchi Han, Orlando Simonetti, Qian Tao

A physics-driven method, Reverse Imaging, enables cardiac MRI segmentation models to generalize across diverse imaging sequences by inferring underlying tissue spin properties from qualitative images. This approach, particularly when used for data augmentation during training, substantially improves segmentation accuracy, achieving Dice scores up to 91.6% on MOLLI and 96.0% on Device datasets with previously unseen contrasts.

Key insights

  • The fundamental 'content' of any MRI image can be understood as underlying tissue spin properties (PD, T1, T2), with image contrast ('style') being a manifestation of these properties governed by MRI physics.
  • Generative diffusion models can effectively serve as powerful priors to regularize the ill-posed inverse problem of inferring physical spin properties from single qualitative MR images.
  • By operating in the interpretable spin property space, the method enables robust zero-shot generalization, synthesizing diverse contrasts for data augmentation and significantly improving segmentation without needing target-domain data for adaptation.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.