Pathology-Foundation-Model-Benchmark

Benchmarking the ability of pathology foundation models to do cell type classification

This repository corresponds to the benchmarking code of the MICCAI 2025 paper:

Lüscher, J., Koreuber, N., Franzen, J., Reith, F.H., Winklmayr, C., Baumann, E., Schürch, C.M., Kainmüller, D. and Rumberger, J.L., 2025, September. Pathocellbench: A comprehensive benchmark for cell phenotyping. In International Conference on Medical Image Computing and Computer-Assisted Intervention (pp. 411-420). Cham: Springer Nature Switzerland.; https://link.springer.com/chapter/10.1007/978-3-032-04981-0_39

Assumes a .env file with the following variables:

• HF_TOKEN: Huggingface API token to download models with access restrictions

Installation

install the project with:

conda create -n <myenv> python=3.10 pip
conda activate <myenv>
pip install -e .

Data

The data can be downloaded via huggingface: https://huggingface.co/datasets/Kainmueller-Lab/phenobench

The dataloaders can be found in the following repository, which is already included in this repository as a dependency: https://github.com/Kainmueller-Lab/Bio-Image-Datasets

GPU requirements

This project is designed to run training and large-model evaluation on NVIDIA GPUs and HPC clusters. Minimal requirements and recommendations:

• CUDA / PyTorch: Built and tested with CUDA 11.8-compatible PyTorch (see pyproject.toml supplemental wheel for cu118). Install a PyTorch build that matches your system's CUDA runtime (CUDA 11.8 recommended).
• NVIDIA drivers & libraries: NVIDIA driver compatible with CUDA 11.8, cuDNN and NCCL installed on the system (or available via container image).

Running the Benchmark

Once you have downloaded the dataset, and installed the required packages, the code can be run with the following command:

Slurm

sbatch launch_train_slurm.sh <pathg_to_config> <model_name>

Grid Engine

qsub launch_train.sh <pathg_to_config> <model_name>

Example: sbatch launch_train_slurm.sh configs/dpt_config.yaml uni2

Cell type classes

The cell type classes are merged as follows:

Cell Type	Category
background	Background
tumor cells	Tumor cells
B cells	B cells
CD11b+ monocytes	Macrophages/Monocytes
CD11b+CD68+ macrophages	Macrophages/Monocytes
CD3+ T cells	T cells
CD4+ T cells	T cells
CD8+ T cells	T cells
CD4+ T cells CD45RO+	T cells
CD4+ T cells GATA3+	T cells
Tregs	T cells
CD163+ macrophages	Macrophages/Monocytes
CD68+ macrophages	Macrophages/Monocytes
CD68+ macrophages GzmB+	Macrophages/Monocytes
CD68+CD163+ macrophages	Macrophages/Monocytes
CD11c+ DCs	Dendritic cells
NK cells	NK cells
adipocytes	Adipocytes
granulocytes	Granulocytes
nerves	Nerves
plasma cells	Plasma cells
smooth muscle	Smooth muscle
stroma	Stroma
vasculature	Vasculature/Lymphatics
lymphatics	Vasculature/Lymphatics
tumor cells / immune cells	Other cells
undefined	Other cells
immune cells	Other cells
immune cells / vasculature	Other cells

Results

Here are the more detailed result tables:

PhenoCell dataset:

Model	Head	Score
HoverNext	N/A	0.238
musk	SimpleSeg	0.200
titan	SimpleSeg	0.232
phikonv2	SimpleSeg	0.243
provgigapath	SimpleSeg	0.264
virchow2	SimpleSeg	0.286
uni2	SimpleSeg	0.303
titan	UNetR	0.264
phikonv2	UNetR	0.280
provgigapath	UNetR	0.280
uni	UNetR	0.288
virchow2	UNetR	0.295
uni2	UNetR	0.301

Arctique:

Model	Head	Score
HoverNext	HoverNext	0.938
musk	SimpleSeg	0.745
phikonv2	SimpleSeg	0.746
titan	SimpleSeg	0.800
uni	SimpleSeg	0.826
uni2	SimpleSeg	0.826
provgigapath	SimpleSeg	0.830
virchow2	SimpleSeg	0.838
titan	UNetR	0.922
phikonv2	UNetR	0.927
provgigapath	UNetR	0.927
musk	UNetR	0.929
uni	UNetR	0.934
virchow2	UNetR	0.943
uni2	UNetR	0.945

PanNuke:

Model	Head	Score
HoverNext	HoverNext	0.784
musk	SimpleSeg	0.639
titan	SimpleSeg	0.691
phikonv2	SimpleSeg	0.697
uni	SimpleSeg	0.710
provgigapath	SimpleSeg	0.714
virchow2	SimpleSeg	0.717
uni2	SimpleSeg	0.726
musk	UNetR	0.766
titan	UNetR	0.781
uni2	UNetR	0.785
provgigapath	UNetR	0.791
uni	UNetR	0.792
virchow2	UNetR	0.794
phikonv2	UNetR	0.797

Lizard:

Dataset	Model	Score
HoverNext	HoverNext	0.725
musk	SimpleSeg	0.571
titan	SimpleSeg	0.601
phikonv2	SimpleSeg	0.607
uni	SimpleSeg	0.636
provgigapath	SimpleSeg	0.655
uni2	SimpleSeg	0.684
virchow2	SimpleSeg	0.689
titan	UNetR	0.697
phikonv2	UNetR	0.697
provgigapath	UNetR	0.697
uni	UNetR	0.716
musk	UNetR	0.722
virchow2	UNetR	0.735
uni2	UNetR	0.729

Domain shift experiments:

Lizard:

Dataset	Model	Score
HoverNext	HoverNext	0.639
musk	SimpleSeg	0.519
phikonv2	SimpleSeg	0.549
titan	SimpleSeg	0.553
provgigapath	SimpleSeg	0.575
uni	SimpleSeg	0.578
uni2	SimpleSeg	0.608
virchow2	SimpleSeg	0.614
titan	UNetR	0.630
uni	UNetR	0.643
phikonv2	UNetR	0.645
provgigapath	UNetR	0.660
musk	UNetR	0.663
virchow2	UNetR	0.670
uni2	UNetR	0.695

PhenoCell DII vs CLR:

Dataset	Model	Score
HoverNext	HoverNext	0.239
musk	UNetR	0.237
titan	UNetR	0.239
uni	UNetR	0.252
phikonv2	UNetR	0.256
provgigapath	UNetR	0.261
virchow2	UNetR	0.266
uni2	UNetR	0.286

PhenoCell MSI vs MSS:

Dataset	Model	Score
HoverNext	HoverNext	0.230
uni	UNetR	0.233
musk	UNetR	0.234
titan	UNetR	0.243
phikonv2	UNetR	0.263
provgigapath	UNetR	0.265
uni2	UNetR	0.284
virchow2	UNetR	0.288

PhenoCell stage 3 vs stage 4:

Dataset	Model	Score
HoverNext	HoverNext	0.209
virchow2	UNetR	0.222
titan	UNetR	0.222
provgigapath	UNetR	0.223
phikonv2	UNetR	0.230
musk	UNetR	0.232
uni	UNetR	0.233
uni2	UNetR	0.251