Multi-Modal Foundation · MahmoodLab
ABMIL (CONCH v1.5 features, pc108-24k)
Attention-based MIL classifier for slide-level cancer subtyping using CONCH v1.5 features.
What ABMIL (CONCH v1.5 features, pc108-24k) is
ABMIL (CONCH v1.5 features, pc108-24k) is a slide-level multiple-instance learning (MIL) classifier developed by MahmoodLab for medical and biological pathology applications. It is designed to evaluate whole-slide histology images by aggregating patch embeddings into slide-level representations.
In computational pathology, large whole-slide images are routinely split into smaller patches. This model uses an attention-based mechanism to weight and pool pre-extracted patch features from the CONCH v1.5 vision model, enabling pan-cancer classification and tumour subtyping for downstream research.
How it works
The model employs an attention-based multiple-instance learning architecture. Instead of processing raw image pixels directly, it ingests patch embeddings generated by the CONCH v1.5 foundation model, using learned attention weights to pool patch vectors into a single slide-level embedding for classification.
The classifier was trained on the pan-cancer pc108-24k dataset to perform slide-level cancer subtyping and histopathological evaluation.
Key facts
- Architecture
- Attention-based Multiple-Instance Learning (ABMIL)
- Input features
- CONCH v1.5 patch embeddings
- Training data
- pc108-24k pan-cancer dataset
- Release date
- 14 July 2025
What you can use it for
- Slide-level cancer subtyping
- Classifying tumour subtypes from pan-cancer whole-slide histology preparations using pooled CONCH v1.5 patch features.
- Downstream pathology evaluation
- Generating slide-level predictions and visual attention distributions for computational pathology benchmarks.
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
- Trained across a broad pan-cancer cohort using the pc108-24k dataset.
- Employs attention-based pooling to aggregate variable numbers of patch representations per slide.
- Builds directly on specialised histopathology feature representations from CONCH v1.5.
Watch out for
- Requires whole-slide images to be pre-tiled and embedded with the CONCH v1.5 model prior to inference.
- Focuses on slide-level classification rather than fine-grained pixel-level segmentation.
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 (CONCH v1.5 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.
What input format does this model expect?
The model takes feature embeddings extracted from histology image patches using the CONCH v1.5 model, rather than raw image pixels.
What training data was used for the model?
The model was trained on the pc108-24k pan-cancer dataset.
What is the primary pathology task addressed by this model?
It is built for slide-level cancer subtyping and downstream pathology classification tasks.
How to cite it
arXiv:2506.09022
Written from ABMIL (CONCH v1.5 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.