Boosting VLMs for Histopathology Classification

The official implementation of Boosting Vision-Language Models for Histopathology Classification: Predict all at once.

This repo is built on top of TransCLIP.

Introduction
Installation
Usage
Citation
Contact

Introduction

Short abstract

We enhance vision-language models (VLMs) for histopathology by introducing a transductive approach that leverages text-based predictions and patch affinity in its objective funtion. Histo-TransCLIP improves zero-shot classification accuracy without additional labels and is able to manage more than 100,000 patches in seconds.

Visual explanation

Figure 1: VLMs leverage textual descriptions of each class to generate pseudo-labels without any manual annotation. These initial predictions are then refined by Histo-TransCLIP by leveraging the data structure thanks to the Laplacian term as well as the text-based predictions.

Results

Dataset	Method	CLIP	Quilt-B16	Quilt-B32	PLIP	CONCH
SICAP-MIL	Zero-shot	29.85	40.44	35.04	46.84	27.71
	Histo-TransCLIP	24.72	58.49	28.18	53.23	32.58
LC(Lung)	Zero-shot	31.46	43.00	76.24	84.96	84.81
	Histo-TransCLIP	25.62	50.53	93.93	93.80	96.29
SKINCANCER	Zero-shot	4.20	15.38	39.71	22.90	58.53
	Histo-TransCLIP	11.46	33.33	48.80	36.72	66.22
NCT-CRC	Zero-shot	25.39	29.61	53.73	63.17	66.27
	Histo-TransCLIP	39.61	48.40	58.13	77.53	70.36

Installation

Environment

Create a Python environment with your favorite environment manager. For example, with conda:

conda create -y --name TransCLIP python=3.10.0
conda activate TransCLIP
pip3 install -r requirements.txt

And install Pytorch according to your configuration:

pip3 install torch==2.0.1 torchaudio==2.0.2 torchvision==0.15.2

Datasets

Install each dataset follow the instructions in DATASETS.md.

Models

Clone repository of each Medical VLM inside Histo-TransCLIP directory:

cd Histo-TransCLIP
git clone https://github.com/mahmoodlab/CONCH.git
git clone https://github.com/PathologyFoundation/plip.git
git clone https://github.com/mlfoundations/open_clip.git

Download Quilt-1m models from hugginface.

Usage

We present the basic usage to get started with our method. You have to pass the datasets folder path and the pre-computed prototypes folder path. Every script has pre-set parameters but you can change them manually.

Histo VLM + TransCLIP-ZS

TransCLIP-ZS based on the textual embeddings of the zero-shot histological models.

Zero-Shot setting

Here is an example for the NCT dataset, with the Quilt-1m architecture, and seed 1:

python3 main.py --root_path /path/to/datasets/folder --dataset nct --method TransCLIP  --seed 1 --model Quilt

To run the whole experiment, use the following command:

bash ./scripts/transclip_zs.sh /path/to/datasets/folder Quilt

Citations

If you find this repository useful, please consider citing our paper:

@article{zanella2024histo,
  title={Boosting Vision-Language Models for Histopathology Classification: Predict all at once},
  author={Zanella, Maxime and Shakeri, Fereshteh and Huang, Yunshi and Bahig, Houda and Ayed, Ismail Ben},
  journal={arXiv preprint arXiv:2409.01883},
  year={2024}
}

Please also consider citing the original TransCLIP paper:

@article{zanella2024boosting,
  title={Boosting Vision-Language Models with Transduction},
  author={Zanella, Maxime and G{\'e}rin, Beno{\^\i}t and Ayed, Ismail Ben},
  journal={arXiv preprint arXiv:2406.01837},
  year={2024}
}

Contact

For any inquiries, feel free to create an issue or contact us at maxime.zanella@uclouvain.be and fereshteh.shakeri.1@etsmtl.net

FereshteShakeri/Histo-TransCLIP