Pathology Features · MahmoodLab
UNI2-h-features
Precomputed UNI2-h patch features for TCGA, CPTAC, and PANDA slides to support downstream pathology models.
What UNI2-h-features is
UNI2-h-features is a collection of precomputed histological feature representations published by MahmoodLab. It contains patch-level embeddings extracted from whole-slide images across the TCGA, CPTAC, and PANDA datasets.
The dataset addresses the computational burden of processing whole-slide pathology images. By providing pre-extracted visual embeddings generated by the UNI2-h model, it serves as a drop-in resource that eliminates the requirement to execute repetitive forward passes over raw histology slides.
How it was made
Histological slides from TCGA, CPTAC, and PANDA are sampled into image patches of 256×256 pixels at 20× magnification. These image patches are then processed through the UNI2-h self-supervised vision transformer model using PyTorch to output numerical patch embeddings.
Key facts
- Covered cohorts
- TCGA, CPTAC, PANDA
- Patch dimensions
- 256×256 pixels
- Magnification
- 20×
- Extraction model
- UNI2-h (self-supervised vision transformer)
- Framework
- PyTorch
- Release date
- 2025-01-11
What you can use it for
- Downstream pathology modelling
- Training multiple-instance learning or slide-level classification systems directly from pre-extracted patch representations without processing raw slide images.
- Cross-cohort histology evaluation
- Conducting comparative studies or validating computational pathology methods across TCGA, CPTAC, and PANDA.
- Algorithmic benchmarking
- Benchmarking novel slide aggregation and analysis architectures using standardised 20× UNI2-h features.
Explore the data
This dataset is gated
Its authors require their terms to be accepted before the data can be read. Accept them on the original dataset page. The platform does not accept terms on your behalf, so the rows are not shown here.Strengths and limits
Good at
- Removes the substantial computational overhead of running UNI2-h forward passes over large slide collections.
- Provides standardized representations extracted at 256×256 resolution and 20× magnification across three distinct pathology cohorts.
Watch out for
- Embeddings are limited to the specific patch size of 256×256 pixels and 20× magnification.
- Coverage is restricted to the TCGA, CPTAC, and PANDA cohorts.
Questions
Can I look inside UNI2-h-features without downloading it?
Not here: its authors gate it behind terms each user must accept on the original page. Its description and statistics are still shown.
What licence is UNI2-h-features under?
Its authors declare cc-by-nc-nd-4.0. Read the terms on the original page before reusing the data.
Which image cohorts are represented in the dataset?
The features are extracted from the TCGA, CPTAC, and PANDA histopathology cohorts.
What patch size and magnification are used?
Features are derived from patches extracted at 256×256 pixels and 20× magnification.
Which model generated these embeddings?
The embeddings were precomputed using the UNI2-h self-supervised vision transformer model.
Written from UNI2-h-features's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.