PFM-DenseBench Overview

A comprehensive benchmark evaluating Pathology Foundation Models on dense prediction tasks

🤖

17

Foundation Models

📁

18

Datasets

⚙️

5

Adaptation Methods

🧪

1530

Total Experiments

Frozen

LoRA

DoRA

CNN Adapter

Transformer Adapter

🏆

H-Optimus-0 achieves the best average rank across all experiments

⚙️

CNN Adapter provides the highest average mDice across all models

📈

mDice scores range from 0.35 to 0.95 depending on task difficulty

📊 Model Rankings

17 of 17 models

Best mDice score achieved on each dataset

Average rank across all datasets (lower is better)

All values are average ranks per dataset (lower is better). Click column headers to sort.

Rank	Model	Overall
Loading data...

Loading performance data...

Click on a row to visit the model's official page

PFM	Architecture	Pretraining	Params.	Slides	Patches	Data Source	Stain	Released	Publication
Vision-only
PathOrchestra	ViT-L/16	DINOv2	303M	300K	-	In-house	H&E	Mar-25	npj Dig. Med.
UNI	ViT-L/16	DINOv2	303M	100,426	100M	GTEx, In-house	H&E	Aug-23	Nat. Med.
UNI2-h	ViT-H/14	DINOv2	681M	350K	200M	GTEx, In-house	H&EIHC	Jan-25	Nat. Med.
Virchow	ViT-H/14	DINOv2	631M	1,488,550	2.0B	In-house	H&E	Sep-23	Nat. Med.
Virchow2	ViT-H/14	DINOv2	631M	3,134,922	2.0B	In-house	H&EIHC	Aug-24	Preprint
Phikon	ViT-B/16	iBOT	86.4M	6,093	43.4M	TCGA	H&E	Jul-23	Preprint
Phikon-v2	ViT-L/16	DINOv2	303M	58,359	456M	132 Public datasets	H&EIHC	Sep-24	Preprint
Prov-GigaPath	ViT-G/14	DINOv2	1.1B	171K	1.4B	In-house	H&EIHC	May-24	Nature
H-Optimus-0	ViT-G/14	DINOv2	1.1B	500K	-	Bioptimus proprietary	H&E	Aug-24	-
H-Optimus-1	ViT-G/14	DINOv2	1.1B	1M	2.0B	Bioptimus proprietary	H&E	Feb-25	Preprint
Midnight-12k	ViT-G/14	DINOv2	1.1B	12K	-	TCGA	H&E	Nov-24	Preprint
Kaiko	ViT-L/14	DINOv2	304M	-	-	-	H&E	2024	Preprint
Lunit	ViT-S/8	DINO	21.7M	-	-	5 Public datasets	H&E	2022	CVPR
Hibou-L	ViT-L/14	DINOv2	304M	1.1M	1.2B	-	H&E	Sep-24	Preprint
Vision-Language
CONCH	ViT-B/16	CoCa	90.4M	21,442	1.17M	PubMed, EDU	H&EIHC	Jul-23	Nat. Med.
CONCHv1.5	ViT-L/16	CoCa	306M	-	1.26M	-	H&EIHC	Nov-24	-
MUSK	ViT-L/16	BEiT3	675M	33K	1M	PubMed, TCGA	H&E	Jan-25	Nature

Click on a row to visit the dataset's official page (Note that all datasets were run at a fixed resolution of 384 on the MUSK model.)

Dataset	Magnification	Anatomical Region	Resize (ViT-14/16)
BCSS	40x	Breast	952 / 944
CoCaHis	40x	Liver	994 / 992
CoNIC2022	20x	Colon	252 / 256
CoNSeP	40x	Colon	994 / 992
COSAS24	-	Multi-organ	994 / 992
CPM15	20x, 40x	Brain	392 / 384
CPM17	20x, 40x	Multi-organ	490 / 496
CRAG	20x	Colon	994 / 992
EBHI	40x	Colon	224 / 224
GlaS	20x	Colon	420 / 416
Janowczyk	40x	Breast	994 / 992
Kumar (MoNuSeg)	40x	Multi-organ	994 / 992
Lizard	20x	Colon	336 / 336
NuCLS	40x	Breast	252 / 256
PanNuke	40x	Pan-cancer	252 / 256
RINGS	100x	Prostate	994 / 992
TNBC	40x	Breast	504 / 512
WSSS4LUAD	10x	Lung	196 / 192