Reproduce “Reconstruct or Generate: Exploring the Spectrum of Generative Modeling for Cardiac MRI” 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
Niklas Bubeck, Yundi Zhang, Suprosanna Shit, Daniel Rueckert, Jiazhen Pan
Researchers from the Technical University of Munich systematically evaluated contemporary generative models for cardiac MRI, reframing reconstruction and generation as a continuous spectrum influenced by data availability. Their "generative model zoo" benchmarked diffusion and autoregressive architectures, revealing diffusion models excel with sufficient context but risk hallucinations when data is sparse, while autoregressive models like MaskGIT offer more stable perceptual quality across varying degrees of missing information.
Key insights
- Medical image reconstruction and generation exist on a continuous spectrum, not as a binary distinction, with model behavior critically dependent on the amount of available conditioning information.
- Perceptual quality can be optimized at both extremes of data availability (very low or very high masking ratios), while intermediate sparsity levels often lead to degraded perceptual quality and increased hallucinations.
- Diffusion models prioritize raw perceptual quality but risk generating anatomically incorrect structures under ambiguity, whereas autoregressive models offer more stable perceptual consistency across varied data sparsity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.