title	emoji	colorFrom	colorTo	sdk	app_file	pinned
Medical image retrieval using a CLIP model	🩺	red	white	streamlit	app.py	true

MedCLIP: Fine-tuning a CLIP model on the ROCO medical dataset

Summary

This repository contains the code for fine-tuning a CLIP model [Arxiv paper][OpenAI Github Repo] on the ROCO dataset, a dataset made of radiology images and a caption. This work is done as a part of the Flax/Jax community week organized by Hugging Face and Google.

[🤗 Model card] [Streamlit demo]

Demo

You can try a Streamlit demo app that uses this model on 🤗 Spaces. You may have to signup for 🤗 Spaces private beta to access this app (screenshot shown below).

The demo can be run locally in the browser with

streamlit run /home/kaushalya/coding/medclip/app.py

Dataset 🧩

Each image is accompanied by a textual caption. The caption length varies from a few characters (a single word) to 2,000 characters (multiple sentences). During preprocessing we remove all images that has a caption shorter than 10 characters. Training set: 57,780 images with their caption. Validation set: 7,200 Test set: 7,650

[ ] Give an example

Installation 💽

This repo depends on the master branch of Hugging Face - Transformers library. First you need to clone the transformers repository and then install it locally (preferably inside a virtual environment) with pip install -e ".[flax]".

Environment setup in Ubuntu or Windows WSL:

conda create --name medclip python=3.10
conda activate medclip
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

    # Optional step: If you don't like CUDA or cdDNN installed in your system, you may install them inside conda environment.
    conda install -c "nvidia/label/cuda-11.8.0" cuda-nvcc
    # Follow the steps in this article to set LD_LIBRARY_PATH in conda environment https://jrkwon.com/2022/11/22/cuda-and-cudnn-inside-a-conda-env/ 
    # Restart the conda environment
    conda deactivate
    conda activate medclip

pip install nvidia-cudnn-cu11 tensorflow==2.12
# cd to your repoisitory folder, like C:/Development
git clone https://github.com/huggingface/transformers.git
cd transformers
pip install -e ".[flax]"
pip install --upgrade "jax[cuda11_pip]" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html
pip install --upgrade flax streamlit optax pandas matplotlib watchdog tables chardet nvidia-tensorrt ipywidgets accelerate

The Model ⚙️

You can load the pretrained model from the Hugging Face Hub with

from medclip.modeling_hybrid_clip import FlaxHybridCLIP

model = FlaxHybridCLIP.from_pretrained("flax-community/medclip-roco")

Alternatively you can download the model checkpoint from [🤗 Model card].

Training

The model is trained using Flax/JAX on a cloud TPU-v3-8. You can fine-tune a CLIP model implemented in Flax by simply running sh run_medclip. This is the validation loss curve we observed when we trained the model using the run_medclip.sh script.

Limitations 🚨

The current model is capable of identifying higher level features such as the modality of ain image (e.g., if a given radiology image is a PET scan or an ultrasound scan). However it fails at identifying a brain scan from a lung scan. ❗️This model should not be used in a medical setting without further evaluations❗️.

Acknowledgements

Huge thanks to the Hugging Face 🤗 team and Google JAX/Flax team for organizing the community week and letting us use cloud compute for 2 weeks. We specially thank @patil-suraj & @patrickvonplaten for the continued support on Slack and the detailed feedback.

TODO

[ ] Mention more examples

[ ] Evaluation on down-stream tasks

[ ] Zero-shot learning performance

[ ] Merge the demo app