MOSAIC

MOSAIC is a benchmark dataset for evaluating the cross-domain generalizability of pathology foundation models, based on the paper Multimodal LLMs Outperform Pathology Foundation Models in Cross-Domain Histological Similarity. The dataset consists of triplets of H&E tile images. Each triplet has an anchor tile and two candidate tiles. One candidate has the same tissue class as the anchor and the other has a different tissue class. The task is to identify the same-class candidate.

Splits

Split	Same-class candidate	Different-class candidate	Datasets
cross_slide	another slide	the anchor's slide	HER2ST, CCTGS
cross_institution	another institution	the anchor's institution	Camelyon16, BEETLE
within_slide	the anchor's slide	the anchor's slide	HER2ST, CCTGS
within_institution	the anchor's institution	the anchor's institution	Camelyon16, BEETLE

In the cross splits the different-class candidate shares the anchor's slide or institution, so slide or institution similarity points to the wrong candidate. In the within splits all three tiles come from one slide or one institution.

Tissue classes

Dataset	Organ	Classes	Tile size (px)
HER2ST	breast	adipose tissue, breast glands, cancer, connective tissue, immune infiltrate	400 × 400
CCTGS	colon	adipose tissue, lamina propria, lymphovascular invasion, muscularis propria, normal mucosa, tumor	512 × 512
Camelyon16	lymph node	normal, tumor	512 × 512
BEETLE	breast	invasive epithelium, non-invasive epithelium, necrosis	512 × 512

Camelyon16 institutions are Radboud and Utrecht. BEETLE institutions are rumc, nki, sch and tcga.

Fields

Field	Type	Description
triplet_id	string	Unique triplet id
category	string	Split name
dataset	string	HER2ST, CCTGS, Camelyon16 or BEETLE
anchor, candidate_1, candidate_2	image	JPEG tiles
<role>_id	string	Tile id {slide}_x{X}_y{Y}_w{W}_h{H}, the tile's position in the source slide image
<role>_slide	string	Slide id
<role>_institution	string	Institution (Camelyon16 and BEETLE only, empty otherwise)
<role>_class	string	Tissue class
answer	int	1 or 2, the candidate with the anchor's tissue class

<role> is anchor, candidate_1 or candidate_2.

Usage

from datasets import load_dataset

mosaic = load_dataset("davidzhanglab/mosaic")
example = mosaic["cross_slide"][0]
example["anchor"], example["candidate_1"], example["candidate_2"], example["answer"]

With an image encoder f, a triplet is solved when the answer candidate is closer to the anchor than the other candidate, for example ‖f(anchor) − f(candidate_answer)‖² < ‖f(anchor) − f(candidate_other)‖². Accuracy is the fraction of solved triplets, usually reported per dataset and split.

Sources

The tiles are crops of H&E slides from four public datasets. Use of the images is subject to the terms of the original datasets.

• HER2ST: Andersson et al. Spatial deconvolution of HER2-positive breast cancer delineates tumor-associated cell type interactions. Nature Communications, 2021. doi:10.1038/s41467-021-26271-2
• CCTGS: Arslan et al. Colorectal cancer tumor grade segmentation: A new dataset and baseline results. Heliyon, 2025. doi:10.1016/j.heliyon.2025.e42467
• Camelyon16: Ehteshami Bejnordi et al. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. JAMA, 2017. doi:10.1001/jama.2017.14585. Litjens et al. 1399 H&E-stained sentinel lymph node sections of breast cancer patients: the CAMELYON dataset. GigaScience,
  2018. doi:10.1093/gigascience/giy065
• BEETLE: Lems et al. A multicentric dataset for training and benchmarking breast cancer segmentation in H&E slides. arXiv, 2025. doi:10.48550/arXiv.2510.02037

Citation

@article{zhang2026multimodal,
  author        = {Zhang, Yishu and Li, Yun and Zhang, Daiwei},
  title         = {Multimodal {LLMs} Outperform Pathology Foundation Models in Cross-Domain Histological Similarity},
  journal       = {arXiv preprint arXiv:2609.32876},
  year          = {2026}
}

Please also cite the source datasets:

@article{andersson2021spatial,
  author  = {Andersson, Alma and Larsson, Ludvig and Stenbeck, Linnea and Salm{\'e}n, Fredrik and Ehinger, Anna and Wu, Sunny Z. and Al-Eryani, Ghamdan and Roden, Daniel and Swarbrick, Alex and Borg, {\AA}ke and Fris{\'e}n, Jonas and Engblom, Camilla and Lundeberg, Joakim},
  title   = {Spatial deconvolution of {HER2}-positive breast cancer delineates tumor-associated cell type interactions},
  journal = {Nature Communications},
  volume  = {12},
  number  = {1},
  pages   = {6012},
  year    = {2021},
  doi     = {10.1038/s41467-021-26271-2},
  pmid    = {34650042}
}

@article{arslan2025colorectal,
  author  = {Duygu Arslan and Sina Sehlaver and Erce Guder and M. A. Temena and Alper Bahcekapili and Umut Ozdemir and Duriye Ozer Turkay and Gunes Guner and Servet Guresci and C. Sokmensuer and Emre Akbas and A. Acar},
  title   = {Colorectal cancer tumor grade segmentation: A new dataset and baseline results},
  journal = {Heliyon},
  volume  = {11},
  number  = {4},
  year    = {2025},
  doi     = {10.1016/j.heliyon.2025.e42467},
  pmid    = {40061913}
}

@article{camelyon16,
  author  = {Babak Ehteshami Bejnordi and M. Veta and Paul Johannes van Diest and B. van Ginneken and N. Karssemeijer and G. Litjens and J. A. van der Laak and M. Hermsen and Quirine F. Manson and Maschenka C. A. Balkenhol and Oscar G. F. Geessink and Nikolaos Stathonikos and Marcory van Dijk and P. Bult and F. Beca and Andrew H. Beck and Dayong Wang and A. Khosla and Rishab Gargeya and Humayun Irshad and Aoxiao Zhong and Q. Dou and Quanzheng Li and Hao Chen and Huangjing Lin and P. Heng and C. Hass and Elia Bruni and Quincy Wong and U. Halici and Mustafa Ümit Öner and R. Cetin-Atalay and Matt Berseth and V. Khvatkov and A. Vylegzhanin and Oren Z. Kraus and M. Shaban and N. Rajpoot and Ruqayya Awan and K. Sirinukunwattana and Talha Qaiser and Yee-Wah Tsang and David Tellez and Jonas Annuscheit and P. Hufnagl and Mira Valkonen and K. Kartasalo and L. Latonen and P. Ruusuvuori and Kaisa Liimatainen and Shadi Albarqouni and Bharti Mungal and A. George and S. Demirci and N. Navab and Seiryo Watanabe and S. Seno and Y. Takenaka and H. Matsuda and Hady Ahmady Phoulady and V. Kovalev and A. Kalinovsky and V. Liauchuk and G. Bueno and M. Fernández-Carrobles and I. Serrano and O. Deniz and Daniel Racoceanu and Rui Venâncio},
  title   = {Diagnostic Assessment of Deep Learning Algorithms for Detection of Lymph Node Metastases in Women With Breast Cancer},
  journal = {JAMA},
  volume  = {318},
  number  = {22},
  pages   = {2199--2210},
  year    = {2017},
  doi     = {10.1001/jama.2017.14585},
  pmid    = {29234806}
}

@article{litjens2018camelyon,
  author  = {G. Litjens and Péter Bándi and Babak Ehteshami Bejnordi and Oscar G. F. Geessink and Maschenka C. A. Balkenhol and P. Bult and Altuna Halilovic and M. Hermsen and Rob van de Loo and Rob Vogels and Quirine F. Manson and N. Stathonikos and A. Baidoshvili and P. V. van Diest and C. Wauters and Marcory van Dijk and J. A. van der Laak},
  title   = {{1399 H\&E-stained sentinel lymph node sections of breast cancer patients: the CAMELYON dataset}},
  journal = {GigaScience},
  volume  = {7},
  number  = {6},
  pages   = {giy065},
  year    = {2018},
  doi     = {10.1093/gigascience/giy065},
  pmid    = {29860392}
}

@article{lems2025multicentric,
  author        = {Carlijn M. Lems and L. Tessier and J. Bokhorst and M. van Rijthoven and W. Aswolinskiy and Matteo Pozzi and Natalie Klubickova and Suzanne Dintzis and Michela Campora and Maschenka C. A. Balkenhol and Peter Bult and J. Spronck and Thomas Detone and M. Barbareschi and E. Munari and G. Bogina and J. Wesseling and E. Lips and F. Ciompi and Frédérique Meeuwsen and J. A. van der Laak},
  title         = {A multicentric dataset for training and benchmarking breast cancer segmentation in H\&E slides},
  journal       = {arXiv preprint arXiv:2510.02037},
  year          = {2025},
  eprint        = {2510.02037},
  archivePrefix = {arXiv},
  doi           = {10.48550/arXiv.2510.02037}
}