Reproduce “Cell Instance Segmentation via Multi-Task Image-to-Image Schrödinger Bridge” 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

Hayato Inoue, Shota Harada, Shumpei Takezaki, Ryoma Bise

Researchers at Kyushu University developed a multi-task image-to-image Schrödinger Bridge framework that directly generates morphologically plausible cell instance masks, effectively bypassing the need for heuristic post-processing steps. The method achieves competitive or superior segmentation performance on public benchmarks like PanNuke and MoNuSeg, demonstrating improved boundary awareness and robustness without reliance on large-scale pre-training.

Key insights

  • Learning a stochastic transport process between an input image and a joint output of instance masks and reverse distance maps enables the generation of morphologically plausible cell shapes.
  • Explicitly guiding the model with reverse distance maps during multi-task training significantly improves boundary delineation and instance separation, particularly in dense cell regions.
  • A generative approach can intrinsically encode global structural constraints, reducing the need for external heuristic post-processing to correct implausible outputs.

Summary by alphaXiv.

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