MedCLIP icon indicating copy to clipboard operation
MedCLIP copied to clipboard

EMNLP'22 | MedCLIP: Contrastive Learning from Unpaired Medical Images and Texts

MedCLIP

PyPI version Downloads GitHub Repo stars GitHub Repo forks

Wang, Zifeng and Wu, Zhenbang and Agarwal, Dinesh and Sun, Jimeng. (2022). MedCLIP: Contrastive Learning from Unpaired Medical Images and Texts. EMNLP'22.

Paper PDF

Download MedCLIP

Before download MedCLIP, you need to find feasible torch version (with GPU) on https://pytorch.org/get-started/locally/.

Then, download MedCLIP by

pip install git+https://github.com/RyanWangZf/MedCLIP.git

# or

pip install medclip

Three lines to get pretrained MedCLIP models

from medclip import MedCLIPModel, MedCLIPVisionModelViT, MedCLIPVisionModel

# load MedCLIP-ResNet50
model = MedCLIPModel(vision_cls=MedCLIPVisionModel)
model.from_pretrained()

# load MedCLIP-ViT
model = MedCLIPModel(vision_cls=MedCLIPVisionModelViT)
model.from_pretrained()

As simple as using CLIP

from medclip import MedCLIPModel, MedCLIPVisionModelViT
from medclip import MedCLIPProcessor
from PIL import Image

# prepare for the demo image and texts
processor = MedCLIPProcessor()
image = Image.open('./example_data/view1_frontal.jpg')
inputs = processor(
    text=["lungs remain severely hyperinflated with upper lobe emphysema", 
        "opacity left costophrenic angle is new since prior exam ___ represent some loculated fluid cavitation unlikely"], 
    images=image, 
    return_tensors="pt", 
    padding=True
    )

# pass to MedCLIP model
model = MedCLIPModel(vision_cls=MedCLIPVisionModelViT)
model.from_pretrained()
model.cuda()
outputs = model(**inputs)
print(outputs.keys())
# dict_keys(['img_embeds', 'text_embeds', 'logits', 'loss_value', 'logits_per_text'])

MedCLIP for Prompt-based Classification

from medclip import MedCLIPModel, MedCLIPVisionModelViT
from medclip import MedCLIPProcessor
from medclip import PromptClassifier

processor = MedCLIPProcessor()
model = MedCLIPModel(vision_cls=MedCLIPVisionModelViT)
model.from_pretrained()
clf = PromptClassifier(model, ensemble=True)
clf.cuda()

# prepare input image
from PIL import Image
image = Image.open('./example_data/view1_frontal.jpg')
inputs = processor(images=image, return_tensors="pt")

# prepare input prompt texts
from medclip.prompts import generate_chexpert_class_prompts, process_class_prompts
cls_prompts = process_class_prompts(generate_chexpert_class_prompts(n=10))
inputs['prompt_inputs'] = cls_prompts

# make classification
output = clf(**inputs)
print(output)
# {'logits': tensor([[0.5154, 0.4119, 0.2831, 0.2441, 0.4588]], device='cuda:0',
#       grad_fn=<StackBackward0>), 'class_names': ['Atelectasis', 'Cardiomegaly', 'Consolidation', 'Edema', 'Pleural Effusion']}

How to Get Sentence-level Semantic Labels

You can refer to https://github.com/stanfordmlgroup/chexpert-labeler where wonderful information extraction tools are offered!