Histopathology vision-language foundation model that turns slides or multi-slide cases into one composite image for retrieval and report generation.
No providers recorded yet. Browse all providers
A pathologist reading a case scans each slide at low power, zooms into the regions that look abnormal, and integrates what they see across every slide from the patient before writing a report. Most pathology foundation models work differently: they cut a whole-slide image into thousands of small tiles, encode each tile on its own, and then aggregate the tile features into a slide-level vector. That pipeline is expensive, separates local morphology from its surrounding tissue context, and treats the single slide rather than the patient case as the unit of analysis.
WILSON, developed by H.R. Tizhoosh's KIMIA Lab and collaborating departments at Mayo Clinic and posted to arXiv in September 2026, replaces the tile-and-aggregate pipeline with a multi-magnification composite. Each slide, or each multi-slide case, is summarized as an 8 × 8 grid of 256-pixel tiles forming one 2048 × 2048 image. Every row is one selected tissue location: a 1.25× context view followed by three 5× and four 20× views of the same region. A single convolutional encoder reads the whole composite in one forward pass, and it is aligned to pathology-report text, so the resulting embedding supports classification by retrieval, image-text search and caption generation.
The vision encoder is a ConvNeXt-Base backbone (87.6M parameters), initialized from DINOv3 natural-image weights, followed by three stride-2 extension blocks (46.2M parameters) that yield a 3,072-dimensional slide embedding. The full encoder has 133.7M parameters and costs about 2.6 TFLOPs per composite. It was pretrained with DINO-style self-distillation, first on single tiles and then on four-tile multi-magnification composites. Training data came from Mayo189K, a diversity-sampled set of 189,291 Mayo Clinic slides drawn from an archive of more than 10 million, covering 42 organs and 829 diagnostic entities. Gemini 2.5 Pro turned each report into three paraphrased captions, giving about 5.28 million image-caption pairs. Four variants share the backbone: dense WILSON, distilled with a SigLIP-style loss from frozen Gemini text embeddings; sparse WILSON, trained on a 991-term keyword vocabulary; WILSON-CLIP, aligned with a PathologyBERT text encoder; and WILSON-CoCa, a captioning model built on the dense encoder.
Without task-specific training, WILSON reached a mean macro-F1 of 0.52 against 0.38 for the case-level model MOOZY in case-level retrieval. In slide-level retrieval it matched TITAN, PRISM and Prov-GigaPath while using 272- to 2,155-fold less compute. Mean recall@1 for image-text retrieval was 75.6% versus 58.1% for PRISM. End-to-end fine-tuning on 508 triple-negative breast cancer cases raised macro-F1 by 0.16 for histologic subtyping and 0.11 for stromal TIL grading. On 1,000 held-out Mayo captions, WILSON-CoCa scored ROUGE-L 0.256, compared with 0.119 for PRISM and 0.167 for PRISM2. The authors note that this internal comparison also reflects WILSON's adaptation to Mayo reporting style.
WILSON is designed for slide and case retrieval in large clinical archives, for search that links pathology text to matching slides, and for drafting diagnostic descriptions that a pathologist then reviews. Its small encoder and single-pass inference suit institutions without the compute for billion-parameter tile pipelines. Groups with only a few hundred labeled cases for a niche task, such as a rare subtype or an institution-specific scanner, can fine-tune the full model rather than training only an aggregator on frozen features.
WILSON argues that representation design, not just scale, drives pathology foundation models: a 133.7M-parameter encoder trained on fewer than 200,000 slides competes with models up to 9.4 times larger. The limitations are concrete. Training data come from a single institution, the 8 × 8 composite layout is one untuned instance of a wider design space, and generated captions contain errors that require pathologist review. The work is a preprint that has not been peer reviewed. The authors state that the model will be released for research download on publication, subject to Mayo Clinic ethics clearance. No code or weights were public at posting, and the Mayo training data are available only as derived annotations under a data-use agreement.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.