🛰️ Daily AI Frontier
‹ back to 2026-08-23

CellPath-Bench: A Multidimensional Benchmark for Whole-Slide Cellular Representations in Pathology Foundation Models

arXiv cs.AI Medical/Healthcare AI Bokai Zhao, Yiyang Zhang, Hanqing Chao, Yawei Ma, Long Bai, Tai Ma, Minfeng Xu, Ming Song, Tianzi Jiang 2026-08-21

TL;DR - CellPath-Bench is a cellular-resolution benchmark for evaluating how well frozen pathology foundation models encode cell-type information in whole-slide images. It enables standardized comparison of both within-section decodability and transfer across tissue sections, datasets, and organs.

  • Uses 25 spatially aligned H&E–Xenium tissue sections spanning 11 organs and more than 7 million cells.
  • Samples frozen whole-slide feature maps at registered nuclear coordinates and evaluates them with standardized multiclass linear probes.
  • Introduces Cell Representation Advantage for comparing nucleus-anchored features with patch-level pooling, and Cell Representation Transferability for measuring cross-domain generalization.
  • Evaluates 30 pathology-specific and general-purpose foundation models across 304,920 runs, revealing distinct model-dependent capability profiles.

view merged work →