WILSON - a pathology foundation model framework for patient-level analysis and diagnostic text generation
Saghir Alfasly, Wataru Uegami, Sobhan Hemati, Wenchao Han, Xiaojia Tang, Kevin Thompson, Daniel Stone, Ghazal Alabtah, et al.
q-bio.QM · cs.AI · cs.CV · cs.LG · eess.IV
Quick verdict
Pathology FM using multi-magnification composite images matches models 9.4× larger at ~1000× less compute; promising but zero-shot, single-institution data.
Summary
WILSON is a vision-language pathology foundation model that replaces the dominant tile-encode-then-aggregate paradigm with a single multi-magnification composite image: an 8×8 grid of 256-pixel tiles sampled by abnormality scoring, forming one 2048×2048 input per slide or multi-slide case. Trained on ~189k Mayo Clinic WSIs (42 organs, ~830 diagnostic entities) with Gemini-generated report captions as supervision, its 133.7M-parameter ConvNeXt encoder was evaluated zero-shot by retrieval on internal and external cohorts. Without task-specific training it exceeded the case-level model MOOZY on all internal cohorts (mean macro-F1 0.52 vs 0.38) and matched WSI-level models up to 9.4-fold larger (PRISM2: 1,251M vs 133.7M parameters) at 272–2,155-fold lower per-WSI compute. End-to-end fine-tuning on 508 triple-negative breast cancer cases raised histologic subtyping macro-F1 by 0.156 and sTIL grading by 0.106. Cross-modal retrieval averaged 75.6% recall@1 versus PRISM's 58.1% across six cohorts, and its CoCa-style captioner led PRISM/PRISM2 on internal and most external caption metrics. Limitations: single-institution training data, modest absolute zero-shot scores, no statistical significance testing, LLM-generated captions serving as both supervision and internal evaluation references, off-the-shelf (not co-adapted) generative baselines, and no stated code or weight release.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 0 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Multi-magnification composite images preserve diagnostically relevant WSI information, matching Yottixel tile selection within 0.05 macro-F1 when both are encoded by the same frozen UNI encoder (Table A1).
- Without task-specific training, WILSON exceeds the case-level model MOOZY on all internal cohorts (mean macro-F1 0.52 vs 0.38) and matches WSI-level foundation models up to 9.4× larger at 272–2,155-fold lower per-WSI compute.
- End-to-end fine-tuning on 508 triple-negative breast cancer WSIs improves histologic subtyping and sTIL grading by +0.156 and +0.106 macro-F1 over the pretrained model.
- WILSON achieves mean cross-modal retrieval recall@1 of 75.6% versus PRISM's 58.1% and generates captions closer to report-derived references than PRISM/PRISM2 on the internal cohort and most external comparisons.
Strengths
Representation claim decoupled from model training
Table A1 encodes composites and Yottixel-selected tiles with the same frozen UNI encoder: macro-F1 differs by ≤0.05 across four benchmarks (composite better on three), directly supporting the preservation claim independent of WILSON's own training.
Breadth of evaluation
Case-level retrieval including PAM50 on external CPTAC-BRCA, WSI-level retrieval, cross-modal retrieval on six cohorts, caption generation, and end-to-end fine-tuning on MayoTNBC — spanning internal Mayo data and public TCGA/CPTAC/HistAI.
Unusually candid comparison caveats
Transparent compute accounting
Leakage-aware evaluation protocol
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
WILSON (Mayo Clinic KIMIA Lab; arXiv 2609.25123) proposes that whole-slide images (WSIs) and multi-slide cases be represented as a single multi-magnification composite — an 8×8 grid of 256-px tiles rendered at 2.5×/10×/20× into one 2048×2048 input — replacing the tile-encode-then-aggregate paradigm of UNI/GigaPath-style foundation models. Five claims are advanced: (1) composites preserve diagnostic information (frozen-UNI comparison vs Yottixel tile selection, |Δ| ≤ 0.05 macro-F1 on four benchmarks, Table A1); (2) zero-shot WILSON beats the case-level model MOOZY on all internal cohorts (mean macro-F1 0.52 vs 0.38; MayoCaseSkin 0.66 vs 0.30, §2.2); (3) it matches WSI-level models up to 9.4× larger (PRISM2: 1,251M vs 133.7M parameters) at 272–2,155-fold lower per-WSI FLOPs (§2.7); (4) end-to-end fine-tuning on 508 TNBC WSIs yields +0.156/+0.106 macro-F1 on histologic subtyping/sTIL grading (§2.4); (5) superior cross-modal retrieval (mean R@1 75.6% vs PRISM 58.1%, Table A11) and captioning. The thesis is representational: pathology-aligned inputs, not scale, drive performance.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 0 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the WILSON pathology foundation model?
WILSON is a vision-language AI model from Mayo Clinic that converts each whole-slide histopathology image — or all slides of a patient case — into a single 2048×2048 composite image combining selected tissue regions at 2.5×, 10× and 20× magnification. It was trained on roughly 189,000 slides with pathology-report text as supervision, and its visual encoder has 133.7M parameters.
How much cheaper is WILSON than other pathology AI models?
At a common workload of 10,000 image tiles per slide it needs about 2.6 trillion FLOPs versus 703 trillion for TITAN, 4.07 quadrillion for PRISM and 5.56 quadrillion for Prov-GigaPath — roughly 272× to 2,155× less compute — and its encoder is 2.7–9.4× smaller than those baselines.
Does WILSON beat PRISM?
In the paper's zero-shot retrieval tests it does on most cohorts: average recall@1 across six image–text retrieval cohorts was 75.6% versus PRISM's 58.1%, and its report captions scored better on internal and most external metrics. However, the comparisons are zero-shot, training data is single-institution (Mayo Clinic), and PRISM ran off-the-shelf while WILSON's captioner was adapted to the test domain, so the gap is not a controlled architecture comparison.
Is WILSON's code or model publicly available?
The paper text reviewed here does not state any code or model-weight release, and training data come from a private Mayo Clinic archive, so independent replication is currently not possible from public artifacts alone.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures are available only as captions; visual claims (e.g., composite legibility, Fig 4c/A8 error cases) could not be assessed from text.
- Text extraction truncated several sections (parts of §2.5, Methods, and Tables B14–B25), so some protocol details were unverifiable.
- Supplementary Notes C–E are only partially quoted, limiting assessment of LLM prompt design and metric caveats.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.