We are witnessing the emergence of many new feature extractors trained using self-supervised learning on large pathology datasets. This repository aims to provide a comprehensive list of these models, alongside key information about them.

I aim to update this list as new models are released, but please submit a pull request / issue for any models I have missed!

Patch-level models

Notes:

• Models marked with VL indicate language-vision pretraining (others are vision-only)
• Models trained on >100K slides may be considered foundation models and are marked in bold
• # of WSIs, tiles, and patients are reported to 2 significant figures
• INE = ImageNet epochs
• Order is chronological
• Some of these feature extractors have been evaluated in a benchmarking study for whole slide classification here.
• ** means inferred from other numbers provided in the paper, repository, or elsewhere
• ^a CONCH1.5 was released as the patch encoder for TITAN, which is a slide-level foundation model. It is also the patch encoder of THREADS.
• Input size refers to the size of the input image at inference time (in pixels)
• Magnification is the effective magnification at which the model sees tissue during pretraining, accounting for any resizing, cropping, or other augmentations. For example, if patches are obtained at 20x with patch size 1024 but resized to 224 before being fed to the model, the effective magnification is 20x × (224/1024) ≈ 4.4x. We assume the following MPP-to-magnification relationship: 0.5 MPP = 20x, 1 MPP = 10x (i.e., MPP × magnification = 10). Often times, models may use random crops; in this case we estimate based on the parameters of the random crop augmentation.

Slide-level / patient-level models

This table includes models that produce slide-level or patient-level embeddings without supervision.