Giga-SSL slide-level foundation models:

Companion package for the paper "Democratizing computational pathology: optimized Whole Slide Image representations for The Cancer Genome Atlas". This package will allow you to encode Whole Slide Images into their Giga-SSL representation (1 feature vector) with one command line.

You can download here the different Giga-SSL models trained using specific tile encoders (moco, ctranspath, phikon, gigapath, H-Optimus-0) as well as the representations of the slides of the TCGA encoded each of these models.

Models are open-source, open-weights, an still constitute (Sept 2024) the state-of-the-art slide-level encoders (GigaSSL + GigaPath features outperform GigaPath slide encoder on private and public benchmark datasets - blogpost coming soon).

Usage - Encode WSIs

Nothing easier. Just use the main.py script; it will handle the necessary model downloads for you.

python main.py --input /path/to/WSI/folder/ --output tcga_brca/ --tile_encoder_type moco

Note: --input is the path of either a WSI or a folder containing WSI (in .ndpi, .svs or .tif format)

Arguments of main.py

• --input: path of either a WSI or a folder containing WSI (in .ndpi, .svs or .tif format) (required)
• --output: path of the output folder (required)
• --tile_encoder_type: type of tile encoder to use (default: moco, available: moco, ctranspath, phikon, gigapath, H-Optimus-0 )
• --gigassl_type: type of gigassl model (default: scm, available: scm, mlp)
• --N_ensemble: number of WSI views to ensemble (default: 100). Usually, the more the better, but the more, the heavier computationally.
• --store_intermediate: path of the folder where to store intermediate results (default: None). Intermediate results include tissue mask, tiles localizations as well as the used tiles themselves.

Parallel computing

Fill the slurm_encode.sh to launch the encodings on multiple nodes.

Test

To run a simple test, install the basic requirements and run:

bash test_one_slide.sh

It will:

• Downloads a slide from the TCGA-BRCA project (and put it under data_test)
• Run the main.py script on it (using gigassl_type=mlp and tile_encoder_type=moco)

It stores the extracted tiles and thumbnail under ./tmp. Have a look at it!

Note: this test ran successfully in 13.3 sec on CPU using the exact environment described in ./tested_env.txt.

Installation

Basic requirements

We recommend using conda to install the necessary dependencies. If you don't have conda installed, please follow the instructions here. Then, in this repository, run:

conda env create -f requirements.yml && conda activate gigassl

With these requirements you will be able to run gigassl models of type mlp using the moco or phikon encoder. They provide a quick way to test the package, as well as good WSIs embeddings (even if a bit less performant than the ones obtained with scm models).

Requirements to run gigassl models of type scm

scm refers to SparseConvMIL-related models. These are the ones evaluated in the paper.

After having installed the basic requirements, you will need to install the SparseConvNet library, following the instruction on its github repository. In their setup instructions, do not forget to have at least one GPU available when launching bash develop.sh, otherwise the GPU version of the package will not be installed.

Requirements to run gigassl trained on top of CTransPath embeddings

scm and mlp GigaSSL models have been trained on both MoCo (in-house model), Phikon and CTransPath embeddings. Performances using the CTransPath embeddings are better than the ones using MoCo embeddings. As stated in the CTransPath original package, you will have to:

• Download a modified timm library here
• install it: pip install timm-0.5.4.tar

Outputs

The main.py script will output a file features_{tile_encoder_type}_{gigassl_type}.csv in the output folder. It is a dictionary mapping WSI names to their corresponding GigaSSL features.

import numpy as np
embeddings = np.load('features_moco_scm.npy', allow_pickle=True).item()

Finally, the file test_classification.py downloads TCGA embeddings computed using the different models and performs classification tasks on them. It provides the classification implementation used in the article.

Here are the results of 10-fold CV AUCs obtained with the 'scm' model trained on top of either CTranspath or MoCo embeddings. These are classification perfomances of simple logistic regression !

	GigaSSL + Moco	GigaSSL + CTransPath	GigaSSL + PhiKon	GigaSSL + Optimus	GigaSSL + GigaPath
Mean ± Std
brca_moltype (TNBC / luminal)	0.933 ± 0.026	0.942 ± 0.035	0.939 ± 0.039	0.968 ± 0.020	0.961 ± 0.029
lung (LUAD / LUSC)	0.960 ± 0.016	0.974 ± 0.010	0.977 ± 0.011	0.981 ± 0.010	0.980 ± 0.010
brca_mhrd (HRD / HRP)	0.793 ± 0.052	0.809 ± 0.050	0.819 ± 0.039	0.834 ± 0.044	0.837 ± 0.048
brca_histotypes (ductal / lobular)	0.924 ± 0.033	0.946 ± 0.040	0.945 ± 0.032	0.943 ± 0.030	0.938 ± 0.034
kidney (cc / p / ch)	0.983 ± 0.008	0.993 ± 0.004	0.994 ± 0.003	0.992 ± 0.004	0.991 ± 0.005