MONET (Medical cONcept rETriever)

MONET is an image-text foundation model trained on 105,550 dermatological images paired with natural language descriptions from a large collection of medical literature. MONET can accurately annotate concepts across dermatology images as verified by board-certified dermatologists, competitively with supervised models built on previously concept-annotated dermatology datasets of clinical images. MONET enables AI transparency across the entire AI system development pipeline from building inherently interpretable models to dataset and model auditing.

• Paper
• GitHub
• BibTex

Getting started

Install

To install the required packages, run the following bash commands:

# clone project
git clone https://github.com/suinleelab/MONET
cd MONET

# [OPTIONAL] create conda environment
conda create -n MONET python=3.9.15
conda activate MONET

# install PyTorch according to instructions at https://pytorch.org/get-started/ v.1.13.0 was used during development.
# example: conda install pytorch==1.13.0 torchvision==0.14.0 pytorch-cuda=11.7 -c pytorch -c nvidia

# install other required python packages
pip install -r requirements.txt
pip install git+https://github.com/openai/CLIP.git

Initialize model

Using original openai CLIP implementation

import clip

def get_transform(n_px):
    def convert_image_to_rgb(image):
        return image.convert("RGB")
    return T.Compose(
        [
            T.Resize(n_px, interpolation=T.InterpolationMode.BICUBIC),
            T.CenterCrop(n_px),
            convert_image_to_rgb,
            T.ToTensor(),
            T.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),        
        ]
    )

model, preprocess = clip.load("ViT-L/14", device="cuda:0", jit=False), get_transform(n_px=224)
model.load_state_dict(torch.hub.load_state_dict_from_url("https://aimslab.cs.washington.edu/MONET/weight_clip.pt"))
model.eval()

Using huggingface CLIP implementation

from transformers import AutoProcessor, AutoModelForZeroShotImageClassification

processor_hf = AutoProcessor.from_pretrained("chanwkim/monet")
model_hf = AutoModelForZeroShotImageClassification.from_pretrained("chanwkim/monet")
model_hf.to("cuda:0")
model_hf.eval()

Usage

We provide jupyter notebooks to demonstrate how to use MONET for automatic concept annotation and various transparency tasks such as data auditing, model auditing, and inherently interpretable model building.

• Automatic concept annotation: tutorial/automatic_concept_annotation.ipynb
• Data auditing: tutorial/data_auditing.ipynb
• Model auditing: tutorial/model_auditing.ipynb
• Inherently interpretable model building: tutorial/inherently_interpretable_model_building.ipynb

MONET Training data

For code to download and preprocess the training data, please refer to the following scripts:

scripts/preprocess/preprocess_pubmed.sh
scripts/preprocess/preprocess_pdf.sh

Training / Evaluation

Code for preprocessing data and training MONET is available in src folder. Code used for evaluation in our paper is available in experiments folder.

Citation

@article{kim2024transparent,
    title={Transparent medical image AI via an image–text foundation model grounded in
medical literature},
    author={Chanwoo Kim and Soham U. Gadgil and Alex J. DeGrave and Jesutofunmi A. Omiye and Zhuo Ran Cai and Roxana Daneshjou and Su-In Lee},
    journal={Nature Medicine},
    year={2024},
    doi={10.1038/s41591-024-02887-x},
    url={https://doi.org/10.1038/s41591-024-02887-x}    
}