MOSAIC: Towards a Universal Latent Space for Computational Pathology Foundation Models

The Key Idea

Do independently trained pathology foundation models learn the same thing?

Central Hypothesis: Independently trained pathology foundation models learn different coordinate systems over a shared underlying morphological manifold.

If that holds, representations should be alignable, semantic neighbourhoods should survive alignment, features should transfer between models, and a downstream classifier should be able to live inside one shared space.

Method

Benchmark 18 state-of-the-art vision and vision-language pathology encoders across 4,331 TCGA and CPTAC whole-slide images, evaluating similarity, representation alignment, and downstream transferability.

RQ1 · Similarity

How similar are the representation spaces of 18 pathology encoders measured by 7 similarity metrics?

RQ2 · Architecture vs Objective

Does the pretraining objective (Vision SSL vs Vision–Language) drive latent geometry more than architecture?

RQ3 · Alignment

Can the evaluated alignment methods align several models into one shared latent space?

RQ4 · Downstream Impact

Does the aligned space improve downstream MIL classification on 13 morphological, clinical, and molecular tasks?

RQ5 · Cross-model Transfer

Does a classifier trained on one encoder generalise to another encoder's features through the shared space?

MOSAIC-Bench

A rigorous benchmark comprising 4,331 whole-slide images and 13 downstream tasks across 3 organ types.

Cohorts (4,331 Slides Total)

TCGA

2,169 Slides

Breast (BRCA) 1,126

Lung (LUAD / LUSC) 1,043

CPTAC

2,162 Slides

Breast (BRCA) 654

Lung (LUAD) 1,139

Colon (COAD) 369

13 Downstream Tasks

Morphological / Clinical

HIGH-LEVEL

Subtyping and staging tasks that are visually apparent in tissue architecture.

NSCLC subtyping (LUAD vs LUSC)
Breast IDC vs ILC
Breast staging
Lung staging

Molecular (Mutations)

LOW-LEVEL

LUAD: TP53, KRAS, STK11
BRCA: PIK3CA, GATA3, MAP3K1
COAD: TP53, KRAS, PIK3CA

The 18 Foundation Models

A diverse landscape of 12 patch-level encoders and 6 slide-level encoders, covering Vision SSL, Vision–Language, and a supervised control.

Resource links point to primary author releases. An em dash means no official resource is available; third-party model mirrors are intentionally omitted.

12 Patch Encoders

Patch-level representation registry

7 Vision SSL 4 Vision–Language 1 Supervised Control

Twelve patch encoders with embedding width, visual backbone, pretraining objective, family, paper, code, and model links.
Encoder	Embedding	Visual Backbone	Pretraining Objective	Family	Paper	GitHub	Hugging Face
UNI2-h	1,536-D	ViT-H/14 + Register Tokens	DINOv2	Vision SSL
Prov-GigaPath	1,536-D	ViT-g/14	DINOv2	Vision SSL
Virchow	2,560-D	ViT-H/14	DINOv2	Vision SSL
Virchow2	2,560-D	ViT-H/14	DINOv2 + Pathology-Specific Augmentations	Vision SSL
H-optimus-0	1,536-D	ViT-g/14	DINOv2	Vision SSL	—
GPFM	1,024-D	ViT-L/14	Multi-Teacher Expert Distillation	Vision SSL
CTransPath	768-D	CNN + Swin-T Hybrid	SRCL Contrastive Learning	Vision SSL			—
CONCH	512-D	ViT-B/16 + Text Encoder	CoCa: Contrastive + Captioning	Vision–Language
CONCH v1.5	768-D	ViT-L/16, 448px	Vision–Language Pretraining (TITAN Patch Encoder)	Vision–Language	—	—
KEEP	768-D	ViT-L/16	Knowledge-Enhanced Vision–Language Pretraining	Vision–Language
MUSK	1,024-D	BEiT-3 Multimodal Transformer, 384px	Masked Modeling + Vision–Language Alignment	Vision–Language
ResNet50	1,024-D	ResNet-50 through Layer 3	ImageNet-1K Supervised Classification	Control			—

6 Slide Encoders

Six slide encoders with embedding width, source grid, paper, code, and model links.
Encoder	Embedding	Built on Grid	GitHub	Hugging Face
CHIEF	768-D	10x / 256px		—
Madeleine	512-D	10x / 256px
PRISM	1,280-D	20x / 224px	—
GigaPath-slide	768-D	20x / 256px
TITAN	768-D	20x / 512px
Feather	512-D	20x / 512px

Theory & Methods

An overview of the 7 similarity metrics and 5 alignment methods used to compare and align cross-encoder representations.

7 Similarity Metrics

Name	What it measures
Linear CKA	Linear correlation between representation spaces independent of orthogonal transformations.
Kernel CKA	Non-linear correlation capturing higher-order interactions using RBF kernels.
SVCCA	Singular Value Canonical Correlation Analysis (PCA followed by CCA).
PWCCA	Projection Weighted CCA (weights CCA directions by variance explained).
Procrustes	Distance after optimal orthogonal alignment.
Cosine RSA	Representational Similarity Analysis using Cosine distance between samples.
Distance Correlation	Dependence measure robust to non-linear associations.

5 Alignment Methods

Name	What it aligns
GCCA	Generalized Canonical Correlation Analysis aligns multiple spaces to a shared latent representation.
MCCA	Multi-set CCA extends standard CCA to maximize pairwise correlations among multiple representation sets.
Generalized Procrustes	Finds optimal orthogonal transformations to align multiple spaces to a central mean representation.
Joint PCA (jPCA)	Computes a shared low-rank subspace across concatenated features.
Shared Autoencoder	Learns non-linear mappings into a shared bottleneck capable of reconstructing all modalities.

Evidence browser

Results by research question

Explore the released findings without scrolling through a wall of figures. Select a research question, then switch cohort, patch grid, or evaluation view. Every panel is regenerated from the released tables by scripts/render_figures.py.

92.5–93.5% GCCA Recall@1 at 10× / 256px across CPTAC and TCGA

7 / 7 similarity measures decline from 5× to 20×

13 tasks compare best single, concatenated, and MOSAIC features

How similar are pathology encoder spaces?

Patch encoders are compared only within matched coordinate grids; slide encoders can be compared directly across grids.