Reproduce “Latent-to-Latent Flow for Volumetric Stochastic Segmentation” on OpenResearch

About this paper

Omar Todd, Sooha Kim, Raghav Mehta, Katherine Mackay, David Bernstein, Alexandra Taylor, Fabio De Sousa Ribeiro, Ben Glocker

Researchers from Imperial College London developed Latent-to-Latent Flow (L2L-Flow), a method for volumetric stochastic segmentation that models label uncertainty by performing flow matching in a lower-dimensional latent space. This approach achieves up to 14x faster inference compared to full-resolution flow models while delivering clinically competitive performance on 3D medical images.

Key insights

• Direct application of existing 2D stochastic segmentation frameworks to high-resolution 3D medical data often leads to performance degradation and instability, necessitating specific methodological adjustments like a time-shifted noise schedule.
• Operating in a learned lower-dimensional latent space for both image and label representations is crucial for scaling probabilistic segmentation to high-resolution 3D volumes efficiently, enabling faster inference.
• Qualitative evaluation against human expert inter-observer variability is vital for assessing clinical relevance, as purely quantitative metrics (e.g., Generalised Energy Distance) may not fully reflect clinically suitable uncertainty in generated samples.

Summary by alphaXiv.