HistoPLUS

Almost everything a pathology foundation model can say about a tumor microenvironment starts with knowing where the cells are and what kind they are. Cell detection, segmentation, and classification on hematoxylin and eosin slides is therefore the substrate for spatial biomarker work, but the models available for it inherit the limits of the public datasets they were trained on. Those datasets cover a handful of common classes well and rare ones badly, and the resulting models degrade when moved to a scanner, stain protocol, or cancer type they have not seen.

HistoPLUS, from Owkin, addresses both problems with a new curated training corpus rather than a new architecture. It uses the CellViT design, a UNETR-style network built on a vision transformer encoder with three decoder branches for nuclei prediction, horizontal-vertical distance maps, and cell type classification, and pairs it with a pathology-specific encoder instead of a general-purpose one. The training set, HistoTRAIN, was annotated specifically to cover cell populations that public benchmarks neglect: neutrophils, eosinophils, macrophages, plasmocytes, endothelial cells, smooth muscle cells, mitotic figures, and apoptotic bodies alongside the usual cancer cells, lymphocytes, fibroblasts, epithelial cells, and red blood cells.

The result sits in the same lineage as Owkin's other pathology work, including the Phikon slide encoders and the OwkinZero reasoning models, but occupies a different rung: it is the cell-level annotator whose outputs feed the downstream spatial and biomarker analyses those models operate on.

#Key Features

• Thirteen-class pan-cancer taxonomy: A single checkpoint distinguishes cancer cells, lymphocytes, fibroblasts, plasmocytes, eosinophils, neutrophils, macrophages, smooth muscle cells, endothelial cells, red blood cells, epithelial cells, mitotic figures, and apoptotic bodies, unlocking study of seven cell types that public datasets barely represent.
• Pathology-specific encoder: The backbone is H0-mini, a distilled DINO/iBOT encoder derived from H-optimus-0. Domain-pretrained encoders consistently beat general-purpose ones such as SAM in the authors' comparison.
• Small enough to deploy: At 86M parameters the model is roughly five times smaller than the CellViT-SAM-H configuration it outperforms, which matters when inference runs over whole-slide images rather than patches.
• Demonstrated cross-domain transfer: Performance holds on two oncology indications absent from training, indicating the model is not simply fitting the cancer types in its corpus.
• Whole-slide tooling: The released package reads common WSI formats at 20x and 40x through OpenSlide and exposes both a CLI and a Python API, emitting JSON annotations with per-cell coordinates, classes, and confidences.

#Technical Details

HistoTRAIN comprises 108,722 segmented nuclei drawn from 1,415 image regions of 448x448 pixels at 40x magnification, sampled from 739 whole-slide images across six cancer types: bladder urothelial carcinoma, colon adenocarcinoma, lung adenocarcinoma, lung squamous cell carcinoma, mesothelioma, and pancreatic adenocarcinoma. Evaluation uses HistoVAL, an independent set of 530 regions containing 69,108 consensus-annotated nuclei from 248 slides, with per-cell-type metrics computed over 1,000 bootstrap iterations and significance testing.

Across four independent external cohorts, HistoPLUS improves detection quality by 5.2% and overall F1 classification score by 23.7% relative to the prior state of the art, with significant gains on 8 of the 13 cell types. On the two indications held out entirely from training it reaches detection quality of 0.836 on breast cancer, with lymphocyte F1 of 0.799, and 0.805 on ovarian cancer, with cancer-cell F1 of 0.682. The encoder ablation is instructive in its own right: pathology-pretrained backbones, including Phikon, Hibou-B, and H0-mini, outperform general-purpose segmentation encoders at equal or smaller parameter counts.

#Applications

The model is aimed at quantitative tumor microenvironment research: counting and spatially mapping immune, stromal, and vascular populations across cohorts, building cell-composition features for biomarker discovery, and quantifying mitotic activity or apoptosis without manual annotation. Computational pathologists and translational research groups can run it over archival H&E cohorts that were never annotated at cell level, which is the practical bottleneck in most spatial biology studies. The authors are explicit that it is a research tool: it is validated only on H&E and only on human tissue, results require expert review, and performance varies with scanner and staining protocol.

#Impact

By showing that careful curation of a modest, well-balanced annotation set beats scale in this task, HistoPLUS makes a case that cell-level pathology models are limited more by label coverage than by encoder capacity. Its efficiency lowers the compute cost of whole-slide cell analysis at cohort scale, and the seven newly accessible cell types open questions about the microenvironment that were previously out of reach. The work is a preprint and has not been through peer review, and its openness is qualified: code and weights are released under CC BY-NC-ND 4.0, which bars both commercial use and derivative models, and access to the weights requires an approved request on HuggingFace. The training corpus itself has not been released.