GrapHist: Graph Self-Supervised Learning for Histopathology

Sevda Öğüt; Cédric Vincent-Cuaz; Natalia Dubljevic; Carlos Hurtado; Vaishnavi Subramanian; Pascal Frossard; Dorina Thanou

cs.CVFeb 24, 2026

GrapHist: Graph Self-Supervised Learning for Histopathology

Authors: Sevda Öğüt, Cédric Vincent-Cuaz, Natalia Dubljevic, Carlos Hurtado, Vaishnavi Subramanian, Pascal Frossard, Dorina Thanou

Organizations: LTS4, EPFL, Lausanne, 1015, Vaud, Switzerland

Abstract

Self-supervised vision models have achieved notable success in digital pathology. However, their domain-agnostic transformer architectures are not originally designed to account for fundamental biological elements of histopathology images, namely cells and their complex interactions. In this work, we hypothesize that a biologically-informed modeling of tissues as cell graphs offers a more efficient representation learning. Thus, we introduce GrapHist, a novel graph-based self-supervised learning framework for histopathology, which learns generalizable and structurally-informed embeddings that enable diverse downstream tasks. GrapHist integrates masked autoencoders and heterophilic graph neural networks that are explicitly designed to capture the heterogeneity of tumor microenvironments. We pre-train GrapHist on a large collection of 11 million cell graphs derived from breast tissues and evaluate its transferability across in- and out-of-domain benchmarks. Our results show that GrapHist achieves competitive performance compared to its vision-based counterparts in slide-, region-, and cell-level tasks, while requiring four times fewer parameters. It also drastically outperforms fully-supervised graph models on cancer subtyping tasks. Finally, we also release five graph-based digital pathology datasets used in our study at https://huggingface.co/ogutsevda/datasets , establishing the first large-scale graph benchmark in this field. Our code is available at https://github.com/ogutsevda/graphist .

Figures & tables

Figure 1: Pre-processing steps in GrapHist. Individual cells are segmented within a digital pathology image. These cells and their spatial arrangement are then converted into a graph. Each cell is a node with morphology, texture, and intensity descriptors. Edges connect neighboring cells and are weighted by their geographic distance. Best viewed in color.

Figure 2: Pre-training of GrapHist. We add a virtual node that is connected to all other nodes in the input cell graph and adopt a masked autoencoding strategy, where a subset of input node features is randomly masked. A GNN-based encoder–decoder architecture is then trained to recover the original node features, decoding from a re-masked version of the latent embeddings.

	# Slides	# Patches	Avg # Nodes	Avg # Edge s	# Classes
Slide-level
TCGA-BRCA [ 43 ]	998	11 149 500	43.94	119.16	2
Region-level
BACH [ 3 ]	394	14 341	35.05	92.93	4
BRACS [ 5 ]	4493	96 153	46.96	128.09	7
BreakHis [ 39 ]	522	709	17.26	41.43	2

Table 1: Statistics for the graph-based slide-, region-, and cell-level datasets.

	Datasets
Model	WSI	RoI
	TCGA-BRCA	BACH	BRACS	BreakHis
Supervised graph models
ACM-bio	OOM	38.37 $\pm$ 2.16	21.55 $\pm$ 2.62	55.82 $\pm$ 10.33
ACM-UNI	OOM	29.03 $\pm$ 9.92	19.95 $\pm$ 3.27	76.27 $\pm$ 9.23
Self-supervised vision models

Table 2: Test macro F1 scores (%) on WSI and RoI-level breast cancer tumor subtyping tasks. For each dataset and model, we report the best MIL aggregation variant. Best and second-best results are highlighted in bold and underlined, respectively.

Panel of Figure 3 — Figure 3: Kaplan-Meier curves with Cox risk groups.

Model	C-inde x	p-value
DINOv2	0.63	9.80 $\times 10^{-6}$
MAE	0.72	1.54 $\times 10^{-13}$
GrapHist	0.76	1.29 $\times 10^{-15}$

Table 3: Survival analysis on TCGA-BRCA.

Model	# Params (10 6 )	Time (h)	Peak GPU mem (GB)	Mean ms/patch
	Pre-training		Inference
DINOv2	22.01	180	0.380	0.885
MAE	47.58	350	0.309	1.053
GrapHist	9.50	50	0.142	0.221

Table 4: Pre-training and inference efficiency on an NVIDIA A100 GPU. Pre-training time is measured on TCGA-BRCA. Inference metrics are measured with BACH.

Panel of Figure 4 — Figure 4: Quantitative results of sensitivity analysis and patch size robustness.

Model	PanNuke 20 $\times$		PanNuke 40 $\times$		NuCLS
	Breast	PanCancer	Breast	PanCancer	main	super
Supervised graph models
ACM-bio	56.61 $\pm$ 0.31	62.73 $\pm$ 0.02	57.06 $\pm$ 0.33	65.80 $\pm$ 0.11	22.13 $\pm$ 1.59	37.35 $\pm$ 1.30
ACM-UNI	58.05 $\pm$ 0.03	67.63 $\pm$ 0.66	56.81 $\pm$ 0.18	66.99 $\pm$ 0.80	21.68 $\pm$ 1.03	39.46 $\pm$ 1.33
Self-supervised vision models
DINOv2	54.82 $\pm$ 0.48	50.49 $\pm$ 0.37	53.86 $\pm$ 0.54	49.27 $\pm$ 0.42	21.42 $\pm$ 1.49	41.17 $\pm$ 2.26

Table 5: Test macro F1 scores (%) on cell-level type identification tasks for all the evaluated models.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 5 — Figure 5: Class label distributions of slide- and region-level datasets.

Panel of Figure 6 — Figure 6: Class label distributions of cell-level datasets.

Nuclei-Color	Nuclei-Morph	Nuclei-Texture
min intensity R	probability	mean ASM
min intensity G	orientation	mean contrast
min intensity B	axis major length	mean correlation
max intensity R	axis minor length	mean dissimilarity
max intensity G	eccentricity	mean energy
max intensity B	area	mean homogeneity

Appendix

Table 6: The set of cell-level color, morphology, and texture descriptors used as input node features in our graph construction pipeline. R, G, and B stand for red, green, and blue, respectively.

Model	Number of Parameters ( $10^{6}$ )
DINOv2	22.01
MAE	47.58
GrapHist ( $d=512$ )	9.50
GrapHist ( $d=768$ )	21.12
GrapHist ( $d=1024$ )	37.34

Appendix

Table 7: Model size comparison across SSL methods. We denote by

d

the validated embedding dimension for GrapHist models.

Stage	Device	Mean sec/patch
Cell segmentation	GPU	0.127
Cell feature extraction	CPU	0.369
Graph construction	CPU	0.009
Total pre-processing	—	0.505

Appendix

Table 8: Pre-processing runtime.

		Datasets
Model	Aggregation	WSI	RoI
		TCGA-BRCA	BACH	BRACS	BreakHis
Supervised graph models
ACM-bio	ABMIL	—	19.56 $\pm$ 5.32	11.80 $\pm$ 1.76	55.82 $\pm$ 10.33
	add-ABMIL	—	38.37 $\pm$ 2.16	21.55 $\pm$ 2.62	49.30 $\pm$ 2.53
	conj-ABMIL	—	19.63 $\pm$ 3.89	13.89 $\pm$ 4.23	53.61 $\pm$ 10.41

Appendix

Table 9: Test macro F1 scores (%) on WSI and RoI-level breast cancer tumor subtyping tasks for all the evaluated models and aggregation methods. Best and second-best results are highlighted in bold and underlined, respectively.

		Datasets
Model	Aggregation	WSI	RoI
		TCGA-BRCA	BACH	BRACS	BreakHis
Supervised graph models
ACM-bio	ABMIL	—	26.23 $\pm$ 4.41	18.69 $\pm$ 3.05	58.50 $\pm$ 8.11
	add-ABMIL	—	41.00 $\pm$ 1.40	25.16 $\pm$ 2.80	50.06 $\pm$ 2.13
	conj-ABMIL	—	24.68 $\pm$ 4.86	18.27 $\pm$ 3.10	56.89 $\pm$ 7.31

Appendix

Table 10: Test balanced accuracy scores (%) on WSI and RoI-level breast cancer tumor subtyping tasks for all the evaluated models and aggregation methods. Best and second-best results are highlighted in bold and underlined, respectively.

		Datasets
Model	Aggregation	WSI	RoI
		TCGA-BRCA	BACH	BRACS	BreakHis
Supervised graph models
ACM-bio	ABMIL	—	50.21 $\pm$ 5.22	57.13 $\pm$ 3.99	60.76 $\pm$ 8.81
	add-ABMIL	—	66.44 $\pm$ 2.58	63.08 $\pm$ 2.48	49.64 $\pm$ 4.38
	conj-ABMIL	—	46.73 $\pm$ 6.13	55.42 $\pm$ 3.16	61.98 $\pm$ 12.31

Appendix

Table 11: Test AUROC scores (%) on WSI and RoI-level breast cancer tumor subtyping tasks for all the evaluated models and aggregation methods. Best and second-best results are highlighted in bold and underlined, respectively.

		Datasets
Model	Aggregation	WSI	RoI
		TCGA-BRCA	BACH	BRACS	BreakHis
Supervised graph models
ACM-bio	ABMIL	—	30.20 $\pm$ 5.67	19.63 $\pm$ 2.77	67.43 $\pm$ 7.26
	add-ABMIL	—	42.91 $\pm$ 1.49	23.35 $\pm$ 2.35	59.13 $\pm$ 3.98
	conj-ABMIL	—	26.45 $\pm$ 3.75	17.98 $\pm$ 2.10	66.39 $\pm$ 8.75

Appendix

Table 12: Test AUPRC scores (%) on WSI and RoI-level breast cancer tumor subtyping tasks for all the evaluated models and aggregation methods. Best and second-best results are highlighted in bold and underlined, respectively.

Model	PanNuke 20 $\times$		PanNuke 40 $\times$		NuCLS
	Breast	PanCancer	Breast	PanCancer	main	super
Supervised graph models
ACM-bio	56.51 $\pm$ 0.41	61.61 $\pm$ 0.33	56.41 $\pm$ 0.38	65.26 $\pm$ 0.59	24.68 $\pm$ 0.96	39.25 $\pm$ 1.36
ACM-UNI	57.38 $\pm$ 0.01	66.72 $\pm$ 0.41	56.96 $\pm$ 0.19	65.64 $\pm$ 0.87	25.29 $\pm$ 0.31	40.53 $\pm$ 1.01
Self-supervised vision models
DINOv2	56.12 $\pm$ 0.50	60.09 $\pm$ 1.15	55.17 $\pm$ 0.57	58.91 $\pm$ 1.35	22.34 $\pm$ 1.76	41.20 $\pm$ 2.42

Appendix

Table 13: Test balanced accuracy scores (%) on cell-level type identification tasks for all the evaluated models. Best and second-best results are highlighted in bold and underlined, respectively.

Model	PanNuke 20 $\times$		PanNuke 40 $\times$		NuCLS
	Breast	PanCancer	Breast	PanCancer	main	super
Supervised graph models
ACM-bio	81.12 $\pm$ 4.53	91.07 $\pm$ 0.07	82.80 $\pm$ 8.75	91.48 $\pm$ 0.23	75.35 $\pm$ 0.69	79.34 $\pm$ 0.43
ACM-UNI	91.39 $\pm$ 2.35	92.92 $\pm$ 0.21	89.05 $\pm$ 3.28	92.38 $\pm$ 0.09	72.02 $\pm$ 0.73	76.13 $\pm$ 1.16
Self-supervised vision models
DINOv2	87.58 $\pm$ 2.28	84.79 $\pm$ 0.45	87.36 $\pm$ 1.98	84.03 $\pm$ 0.44	63.44 $\pm$ 0.52	68.09 $\pm$ 2.02

Appendix

Table 14: Test AUROC scores (%) on cell-level type identification tasks for all the evaluated models. Best and second-best results are highlighted in bold and underlined, respectively.

Model	PanNuke 20 $\times$		PanNuke 40 $\times$		NuCLS
	Breast	PanCancer	Breast	PanCancer	main	super
Supervised graph models
ACM-bio	61.49 $\pm$ 0.23	67.80 $\pm$ 0.17	62.44 $\pm$ 0.02	70.37 $\pm$ 0.06	28.29 $\pm$ 0.46	47.78 $\pm$ 0.35
ACM-UNI	64.11 $\pm$ 0.12	73.59 $\pm$ 0.66	62.40 $\pm$ 0.06	72.76 $\pm$ 0.43	26.93 $\pm$ 0.44	45.20 $\pm$ 0.31
Self-supervised vision models
DINOv2	59.52 $\pm$ 0.73	55.64 $\pm$ 1.35	58.69 $\pm$ 0.74	54.25 $\pm$ 1.17	24.63 $\pm$ 1.31	43.97 $\pm$ 2.09

Appendix

Table 15: Test AUPRC scores (%) on cell-level type identification tasks for all the evaluated models. Best and second-best results are highlighted in bold and underlined, respectively.

Explore similar work

Learning Where to Focus: Self-Supervised Multi-Scale ViTs for Histopathology
Sep 16, 2026Anabel Stammer, Valay Bundele, Mehran Hosseinzadeh +1Computational Pathology Self-Supervised Visual Representation Learning
Eberhard Karls Universität Tübingen, Germany
DaX: Learning General Pathology Representations Across Scales
Jun 5, 2026Bokai Zhao, Yiyang Zhang, Long Bai +3Pathology Foundation Models Medical Image Representation Learning
DAMO Academy, Alibaba Group · Institute of Automation, Chinese Academy of Sciences · Hupan Lab
DeepHistoViT: An Interpretable Vision Transformer Framework for Histopathological Cancer Classification
Mar 12, 2026Ravi Mosalpuri, Mohammed Abdelsamea, Ahmed Karam EldalyHistopathology Image Classification Computational Pathology

Sep 16, 2026J/K move · Enter open · S save

Learning Where to Focus: Self-Supervised Multi-Scale ViTs for Histopathology

Anabel Stammer, Valay Bundele, Mehran Hosseinzadeh, Hendrik P. A. Lensch

Eberhard Karls Universität Tübingen, Germany

Computational Pathology Self-Supervised Visual Representation Learning Vision Transformer Efficient ViTs

Open paper AskPDF arXiv

Pathologists diagnose diseases by first locating suspicious tissue and then examining it at higher magnification, whereas self-supervised vision transformers (ViTs) allocate the same spatial resolution to every image region despite diagnostic evidence being sparse and spanning multiple biological scales. Recent pathology foundation models have substantially improved representation quality by scaling training data and model capacity, but largely retain uniform tokenization. We instead investigate whether pathology representations can be improved by learning where to allocate spatial resolution during self-supervised learning. To this end, we propose CRAFT (Coarse-to-fine Region-Adaptive Feature Tokenization), a DINO-based framework that learns image-dependent mixed-scale representations by using self-supervised attention to selectively refine informative regions while preserving coarse context, together with a symmetric cross-scale regularization objective that encourages complementary coarse and fine representations. Across CAMELYON16, TCGA-Lung subtype classification, and TCGA-LUAD survival prediction, CRAFT consistently outperforms comparable-scale self-supervised methods while requiring lower inference computation. Despite using only a compact 22M parameter backbone trained on comparatively small pathology datasets, CRAFT remains competitive with, and often surpasses, substantially larger pathology foundation models.

Abstract

Explore similar work

Learning Where to Focus: Self-Supervised Multi-Scale ViTs for Histopathology

DaX: Learning General Pathology Representations Across Scales

DeepHistoViT: An Interpretable Vision Transformer Framework for Histopathological Cancer Classification