Kyushu University's Schrödinger Bridge Beats CellViT at Cell Segmentation Without Post-Processing

Cell segmentation pipelines have quietly become Rube Goldberg machines. A convolutional network predicts a probability map, a distance map, or a vector field, then a stack of watershed transforms, morphological operations, and hand-tuned thresholds tries to tease apart cells that are touching. A new paper from Kyushu University tosses out most of that plumbing and treats the whole problem as a Schrödinger Bridge, learning a direct diffusion path from a histology image to its instance mask.

The result, called Multi-Task Image-to-Image Schrödinger Bridge, beats CellViT on PanNuke without using SAM pretraining and without any post-processing pipeline. A generic generative model, trained from an ImageNet checkpoint, edges out a Vision Transformer specifically built on top of Meta's Segment Anything backbone.

Why post-processing is the weak link

In densely packed microscopy images, the boundary between two touching cells is often a single pixel wide, and ground-truth masks label that boundary as background. Most existing frameworks lean on post-processing to separate adjacent cells: a CNN predicts intermediate representations like probability maps, distance maps, or vector fields, and individual instances are recovered through watershed transforms, clustering, or morphological operations.

That works, but it has two problems. Such pipelines often depend on dataset characteristics and hyperparameter choices, which limits robustness and generalizability. And because the network output is deterministic, nothing prevents it from producing morphologically absurd shapes like merged cells or spurious fragments, which the downstream cleanup then has to fix.