Multi-Modal Foundation · MahmoodLab
ABMIL (UNI features, pc108-24k)
Slide-level attention-based classifier aggregating UNI patch features across pan-cancer histology images.
What ABMIL (UNI features, pc108-24k) is
ABMIL (UNI features, pc108-24k) is a computational pathology model created by MahmoodLab for medical and biological applications. It functions as a slide-level classifier using attention-based multiple instance learning to evaluate whole-slide tissue images.
In digital pathology, microscopic slides are often too large to process in a single pass and are divided into smaller patches. This model addresses that challenge by providing slide-level aggregation over patch features derived from the original UNI patch encoder, combining local visual information into a unified slide-level representation.
How it works
The model employs an attention-based multiple instance learning (ABMIL) architecture designed to process sets of patch embeddings. It takes feature representations produced by the UNI patch encoder as input and computes attention weights across patches to aggregate them into slide-level classifications. Training was conducted using the pc108-24k pan-cancer dataset.
Key facts
- Architecture
- Attention-based multiple instance learning (ABMIL)
- Task
- Slide-level multiple instance learning classification
- Input features
- UNI patch encoder embeddings
- Training data
- pc108-24k pan-cancer dataset
What you can use it for
- Whole-slide classification
- Categorises whole-slide histology specimens by aggregating patch-level embeddings into a single slide assessment.
- Slide-level feature aggregation
- Combines patch representations generated by the UNI foundation model into unified slide-level vectors using attention mechanisms.
- Pan-cancer pathology research
- Analyses digital histology slides across cancer types using representations derived from the pc108-24k training cohort.
Run it here
We list this model but cannot run it yet: excluded (gated): Gated and CC-BY-NC-ND-4.0..
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Directly aggregates representations from the UNI patch encoder for whole-slide tasks.
- Uses attention weighting to highlight relevant patches during slide-level classification.
Watch out for
- Operates exclusively on pre-extracted UNI feature embeddings rather than raw whole-slide images.
- Requires custom PyTorch execution code to load and run the architecture.
Inputs and outputs
You provide
Image
PNG, JPEG or TIFF. Sent to the model as-is.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · cc-by-nc-nd-4.0
This model is gated by its authors — its licence has to be accepted on the source repository before it will run.
Questions
Can I use ABMIL (UNI features, pc108-24k) commercially?
Check the licence before commercial use. This model is gated by its authors — its licence has to be accepted on the source repository before it will run.
Can this model process raw histology images directly?
No. The model requires pre-extracted feature representations generated by the UNI patch encoder as its input.
What dataset was used to train this model?
The classifier was trained on pc108-24k, a pan-cancer pathology dataset.
What is the primary output of the model?
It generates slide-level classifications and aggregated slide representations from sets of patch-level embeddings.
How to cite it
arXiv:2506.09022
Written from ABMIL (UNI features, pc108-24k)'s own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.