MOOZY

A Patient-First Foundation Model for Computational Pathology

Yousef Kotp^1,2, Vincent Quoc-Huy Trinh^3,4, Christopher Pal^2,5,6, Mahdi S. Hosseini^1,2,7

¹CSSE, Concordia University, Montréal, Canada ²Mila – Québec AI Institute, Montréal, Canada ³CHUM, Université de Montréal, Montréal, Canada ⁴IRIC, Université de Montréal, Montréal, Canada ⁵Université de Montréal, Montréal, Canada ⁶Canada CIFAR Chair, Polytechnique Montréal, Montréal, Canada ⁷Dept. of Pathology, McGill University, Montréal, Canada

🏆 Accepted at ECCV 2026

📄 Paper 💻 Code 🤗 Model 📦 PyPI

85.77M parameters

77,134 public slides

333 tasks from 56 datasets

23 anatomical sites

Quick Start

# install
pip install moozy

# encode from pre-computed H5 feature files
moozy encode slide_1.h5 slide_2.h5 -o case_embedding.h5

# encode from raw whole-slide images (requires AtlasPatch, SAM2, and OpenSlide)
moozy encode slide_1.svs slide_2.svs -o case_embedding.h5 --target_mag 20

Checkpoint and task definitions are downloaded automatically from HuggingFace on first use.

Abstract

Computational pathology needs whole-slide image (WSI) foundation models that transfer across diverse clinical tasks, yet current approaches remain largely slide-centric, often depend on private data and expensive paired-report supervision, and do not explicitly model relationships among multiple slides from the same patient. We present MOOZY, a patient-first pathology foundation model in which the patient case, not the individual slide, is the core unit of representation. MOOZY explicitly models dependencies across all slides from the same patient via a case transformer during pretraining, combining multi-stage self-supervision with scaled low-cost task supervision. In Stage 1, we pretrain a vision-only slide encoder on 77,134 public slide feature grids using masked self-distillation. In Stage 2, we align these representations with clinical semantics using a case transformer and multi-task supervision over 333 tasks from 56 public datasets, including 205 classification and 128 survival tasks across four endpoints. Across sixteen held-out tasks, MOOZY improves macro weighted F1, balanced accuracy, and macro weighted ROC-AUC relative to PRISM by +4.19%, +7.93%, and +6.95%, respectively. MOOZY is also parameter efficient with 85.77M parameters, 14× smaller than GigaPath. These results suggest that patient-level pretraining yields transferable embeddings, providing a path toward scalable patient-first histopathology foundation models.

Two-Stage Training Pipeline

Stage 1 (top): A frozen patch encoder extracts per-patch features arranged into a spatial grid. Multi-scale crops are sampled with spatial augmentations and block-based masking. A student slide encoder and EMA teacher are jointly trained via CLS-level self-distillation and masked patch prediction. Stage 2 (bottom): The pretrained slide encoder produces per-slide embeddings; a case transformer aggregates them into a unified case embedding, routed to task-specific classification and survival heads.

Training Data Scale

MOOZY is trained entirely on public data. Stage 1 uses 77,134 slide feature grids (53,286 at 20× and 23,848 at 40×) extracted from ~1.67 billion patches across ~31.8 TB of raw WSI data. Stage 2 uses 41,089 supervised cases (45,179 unique whole-slide images) across 333 tasks from 56 datasets — all 32 TCGA cohorts, all 10 CPTAC cohorts, REG, BC-Therapy, BRACS, CAMELYON17, DHMC Kidney, DHMC LUAD, EBRAINS, IMP Colorectum, IMP Cervix, MBC, MUT-HET-RCC, NADT Prostate, NAT-BRCA, and PANDA. Supervision covers 205 classification and 128 survival tasks across four endpoints (OS, DSS, DFI, PFI) and 23 anatomical sites.

MLP-probe weighted F1 radar comparison across slide encoders

(a) MLP-probe weighted F1 across sixteen held-out tasks. Brackets show [min–max] per task (center = min, outer ring = max).

Linear-probe weighted F1 radar comparison across slide encoders

(b) Linear-probe weighted F1 across the same sixteen held-out tasks.

Macro-averaged weighted F1 versus total parameter count

(c) Macro-averaged weighted F1 (MLP probe) vs. total parameter count (log scale). Bubble size indicates total parameters.

Frozen-feature MLP probe on sixteen held-out tasks. Bold = best, underline = second best. Mean ± std over 5 folds.

Task	Metric	CHIEF	GigaPath	PRISM	Madeleine	TITAN	MOOZY
Residual Cancer Burden	F1	0.46	0.45	0.46	0.51	0.43	0.56
	AUC	0.60	0.55	0.58	0.63	0.58	0.74
	Bal. Acc	0.44	0.40	0.43	0.48	0.38	0.51
TP53 Mutation	F1	0.82	0.76	0.85	0.84	0.87	0.87
	AUC	0.81	0.76	0.85	0.85	0.91	0.86
	Bal. Acc	0.83	0.76	0.84	0.84	0.88	0.86
BAP1 Mutation	F1	0.86	0.84	0.80	0.85	0.84	0.89
	AUC	0.75	0.63	0.71	0.78	0.82	0.79
	Bal. Acc	0.75	0.66	0.66	0.75	0.75	0.78
ACVR2A Mutation	F1	0.89	0.80	0.85	0.89	0.87	0.91
	AUC	0.80	0.74	0.83	0.76	0.79	0.91
	Bal. Acc	0.80	0.65	0.81	0.81	0.76	0.90
Histologic Grade	F1	0.71	0.77	0.73	0.75	0.73	0.78
	AUC	0.71	0.77	0.67	0.74	0.71	0.75
	Bal. Acc	0.73	0.77	0.73	0.74	0.73	0.77
KRAS Mutation	F1	0.77	0.77	0.72	0.81	0.80	0.85
	AUC	0.76	0.72	0.61	0.70	0.80	0.80
	Bal. Acc	0.74	0.76	0.63	0.77	0.81	0.79
IDH Status	F1	0.92	0.94	0.91	0.92	0.94	0.97
	AUC	0.96	0.97	0.95	0.96	0.97	0.99
	Bal. Acc	0.92	0.94	0.91	0.91	0.94	0.97
Treatment Response	F1	0.53	0.51	0.57	0.49	0.49	0.58
	AUC	0.70	0.68	0.69	0.59	0.60	0.68
	Bal. Acc	0.48	0.40	0.51	0.35	0.37	0.48
BRCA PAM50 Subtype	F1	0.67	0.68	0.70	0.68	0.72	0.63
	AUC	0.83	0.83	0.85	0.84	0.87	0.80
	Bal. Acc	0.51	0.52	0.57	0.53	0.58	0.50
HNSC mRNA Subtype	F1	0.54	0.60	0.60	0.59	0.62	0.55
	AUC	0.75	0.77	0.78	0.77	0.79	0.72
	Bal. Acc	0.55	0.61	0.59	0.58	0.61	0.54
UCEC Genomic Subtype	F1	0.55	0.57	0.63	0.57	0.66	0.56
	AUC	0.75	0.76	0.81	0.75	0.82	0.74
	Bal. Acc	0.53	0.54	0.60	0.55	0.62	0.52
Synaptophysin Grade	F1	0.78	0.79	0.76	0.81	0.80	0.79
	AUC	0.61	0.74	0.50	0.68	0.64	0.61
	Bal. Acc	0.64	0.69	0.58	0.65	0.68	0.65
RAS/BRAF Status	F1	0.89	0.89	0.84	0.88	0.90	0.94
	AUC	0.68	0.48	0.56	0.33	0.80	0.75
	Bal. Acc	0.68	0.68	0.57	0.66	0.68	0.85
Keratinizing SCC Grade	F1	0.74	0.68	0.72	0.69	0.72	0.73
	AUC	0.75	0.69	0.74	0.72	0.72	0.72
	Bal. Acc	0.73	0.68	0.72	0.69	0.71	0.72
Non-Keratinizing SCC Grade	F1	0.80	0.78	0.75	0.80	0.81	0.77
	AUC	0.66	0.61	0.52	0.66	0.71	0.66
	Bal. Acc	0.79	0.75	0.72	0.79	0.78	0.73
Primary vs. Metastasis	F1	0.92	0.92	0.92	0.93	0.93	0.93
	AUC	0.62	0.60	0.66	0.75	0.77	0.69
	Bal. Acc	0.57	0.58	0.63	0.63	0.64	0.65

Macro-average across sixteen held-out tasks. Each entry averages over five MIL architectures (MeanMIL, ABMIL, CLAM, DSMIL, TransMIL).

Metric	Backbone	UNI v2	Phikon v2	CONCH v1.5	MUSK	MOOZY
F1 (weighted)	0.723	0.722	0.714	0.740	0.720	0.769
ROC-AUC (weighted)	0.707	0.707	0.697	0.720	0.695	0.763
Balanced Acc	0.643	0.637	0.625	0.661	0.637	0.702

Macro-average across sixteen held-out tasks. Each stage and the case aggregator are toggled independently.

Setting	Stage 1	Stage 2	Case Agg.	F1	AUC	Bal. Acc
Stage 1 only	✓	✗	✗	0.743	0.715	0.662
Stage 2 only w/o case agg.	✗	✓	✗	0.743	0.721	0.664
Stage 2 only	✗	✓	✓	0.731	0.697	0.659
MOOZY w/o case agg.	✓	✓	✗	0.749	0.737	0.682
MOOZY	✓	✓	✓	0.769	0.763	0.702

Where Does MOOZY Look?

A board-certified pathologist reviewed attention maps across 20 representative WSIs sampled from eight held-out evaluation cohorts and five encoders. MOOZY achieved the lowest mean semantic gap score (1.00, versus 1.38 for TITAN and 1.75 for PRISM) and near-balanced tumor vs. non-tumor attention (shift 2.63), suggesting broad, diagnostically relevant coverage.

Attention-map comparison on a lung adenocarcinoma slide. MOOZY and TITAN: balanced, comprehensive coverage (shift 3, gap 1). PRISM: balanced shift with moderate gaps (shift 3, gap 2). CHIEF and Madeleine: cancer-biased with frequent semantic gaps (shift 2, gap 3).

Embedding Quality

Dimensionality reduction of slide embeddings from four encoders. MOOZY shows the clearest class separation on cancer-type tasks.

MOOZY

TITAN

Madeleine

PRISM

Citation

@inproceedings{kotp2026moozypatientfirstfoundationmodel,
      title={MOOZY: A Patient-First Foundation Model for Computational Pathology},
      author={Kotp, Yousef and Trinh, Vincent Quoc-Huy and Pal, Christopher and Hosseini, Mahdi S.},
      booktitle={European Conference on Computer Vision (ECCV)},
      year={2026},
      url={https://arxiv.org/abs/2603.27048},
}